杰弗里·C·鲍克

Geoffrey C. Bowker

算法渗透到我们的生活。它们是政治、文化和社会事实，在过去五十年中已成为我们生活各个方面的核心。当然，我们之前就有了它们的前身：无休止的清单、安全协议和行为准则——每一项都旨在让我们脱离自我，将我们的身体、我们的自我与官僚或技术机器（用福柯更好的术语来说，是一套“设备技术”）对齐。官僚主义让我们像机器一样行事，算法则试图让我们变成机器。

Algorithms pervade our lives. They are political, cultural, and social facts that have become central to all parts of our existence over the past fifty years. Certainly, we had their forerunners before: endless checklists, safety protocols, and rules of conduct—each designed to take us out of ourselves and align our bodies, our selves with a bureaucratic or technical machine (in Foucault’s better term, a set “dispositifs techniques”). Bureaucracy makes us act like machines, algorithms seek to make us into machines.

一个必然结果是，如果我们想做基础社会科学，设想新的政治生活形式，我们就需要去行动的地方。我们需要从内部了解算法。它们不是从另一个星球空降下来入侵我们（尽管感觉就像这样）：它们是人类的、会犯错的创造物。这里的困难在于，社会科学家和政治参与者往往并不真正了解技术利害关系，而计算机科学家同样不真正了解社会利害关系。

A corollary is that if we want to do fundamental social science and envision new forms of political life we need to go where the action is. We need to get to know algorithms from the inside. They did not parachute down from another planet to invade us (much as it may feel like this): they are human, fallible creations. The difficulties here are that social scientists and political actors often don’t really understand the technical stakes, and symmetrically the computer scientists don’t really get the social stakes.

这正是这本书如此重要的原因。它是探索算法作为一种新型社会行为体的基础文本。算法是如何被构建为有效的行为体？人类是如何被构建的，以便他们创造出超越人类理解的算法？Jaton 的探索是无畏的：去问题所在的地方，在技术、社会和政治问题中找到它们的主场。当我读这本书时，我总是很高兴，就像读一本好小说一样，不知道接下来会发生什么（冯·诺依曼架构，对新生计算机工程师的测试）——但一旦采取了这些步骤，就会立即感受到一种必然性。

This is precisely why this book is so important. It is a foundational text for exploring algorithms as a new form of social actor. How do algorithms get constructed to be effective actors; how do humans get constructed so that they create algorithms which surpass human understanding? Jaton’s quest here has been fearless: go where the questions are, and locate the technical, social, and political issues on their home ground. As I read this book, I was constantly delighted as when reading a fine novel by not knowing what was going to come next (von Neumann architecture, tests for nascent computer engineers)—but by immediately feeling a sense of inevitability once the steps were taken.

最近我一直在思考这样一个问题：人类与政治经济的运作越来越不相关：我们尽我们所能，但却越来越处于间隙之中。毫无疑问，我们我们正在创造比我们更聪明的机器和比我们更了解我们的算法。这很好。但如果我们将数千年来根深蒂固的价值观融入算法中，我们将创造出多么丰富和美丽的世界？

I’ve been playing with a vision latterly of humans becoming progressively more irrelevant to the operation of our political economy: we do what we can but are increasingly interstitial. There is little doubt that we are creating machines that are more intelligent than we are and algorithms that know us better than we do ourselves. That’s just fine. But how much richer and more beautiful a world we will create if we suffuse our algorithms with our own deeply held values created over thousands of years?

这本书不只是为计算机科学家或科学社会研究学者而写的：它探讨了这个时代人类生存的一些基本问题。它为我们提供了共同设计我们的世界（我们的行星系统、我们的物种、我们的计算机）的工具和概念。

This book is not just for computer scientists or for social studies of science scholars: it speaks to some of the fundamental questions of human existence in this epoch. It provides tools and concepts for us to co-engineer our world (our planetary system, our species, our computers).

开场白！Florian。祝大家阅读愉快。

Chapeau! Florian. Happy reading all.

负面隐形

Negative Invisibilities

此时，人们可能会产生一个问题：为什么解释算法的形成过程很重要？为什么要花时间和精力撰写和阅读有关算法的构成？除了让算法产生的活动变得清晰可见之外，难道没有其他事情可做吗？

Already at this point, a question may arise: Why is it important to account for the formation processes of algorithms? Why spending time and energy writing—and reading—about their constitution? Are there not other things to do than making the activities by which algorithms come into existence visible?

当然。正如 Star 和 Strauss (1999) 所言，有些活动需要暂时保持隐形——也就是说，不被解释——否则这些活动的结果可能会失去一些能力。马戏团就是一个例子：将设计和掌握太阳马戏团的空中飞人表演所需的基础设施和培训实践公开，可能会对表演本身产生负面影响。奇迹、惊喜或魅力可能会被促成表演的脚踏实地和不确定的运作所抵消。在这里，社会学解释会冒着破坏表演的风险；它可能会降低表演的^表演能力。6按照 Star 和 Strauss (1999, 23) 的区分，空中飞人表演的相对隐形性在这种意义上是积极的：它有助于这些马戏团实践的产品，由于没有更好的术语，可以称之为充分。缺乏任何公开可用的信息，账户和保密的存在有助于行为成为一种行为，就像它们有助于公众成为行为的公众一样。在这种非常特殊的情况下，人们可以假设双方都希望相信掌握。

Certainly. As Star and Strauss (1999) have suggested, some activities need to remain provisionally invisible—that is, not accounted for—otherwise the results of these activities may lose some of their capacities. The circus is one example: making publicly visible the infrastructure and training practices required to design and master, say, a Cirque du Soleil trapeze act may negatively affect the act itself. Wonder, surprise, or enchantment would potentially be counteracted by the down-to-earth and uncertain operations that enabled the act. Here, a sociological account would take the risk of spoiling the act; it may lower the act’s capacity to act.⁶ Following the distinction made by Star and Strauss (1999, 23), the relative invisibility of the trapeze act is, in that sense, positive: it helps the product of these circus practices to be, by lack of a better term, adequate. The lack of any publicly available account and the presence of secrecy help the act become an act, just as they help the public become the public of the act. In such a very specific situation, one may assume there is a mutual desire to believe in mastery.

但是，一旦对某些实践的成果产生争议，其充分性就会受到质疑；当某种行动能力受到质疑时，就需要正视其形成方式的不同意见。举例来说，假设同样的太阳马戏团高空秋千表演导致了一场事故。如果对这场事故产生争议，就会有人要求公开导致事故发生的实践。完成这场高空秋千表演所需的实践将从积极的隐形变成消极的隐形：为了使争议的不同方能够进行谈判，就需要对这一表演如何产生进行实证分析。太阳马戏团需要什么来表演这场有争议的表演？哪些元素可以改变来重新调整这个脆弱的组合？简而言之，为了提出妥协方案，为了更好地创作，争论者将受益于对高空秋千实践的实证分析；⁷记录表演者和艺人所珍视和害怕的事物以及他们所依附的事物，可能会让关于他们所制作作品的代理机构的建设性分歧得以展开。

But as soon as there are controversies about the products of some practices, the terms of their adequacy are disputed; when some capacity to act is put into question, disagreements about its formation need to be confronted. Let us, for example, imagine that the same Cirque du Soleil trapeze act leads to an accident. If disputes arise about this accident, there will be requests to make visible the practices that contributed to producing it. From being positively invisible, the practices required to do this trapeze act would become negatively invisible: for the different parties of the dispute to become able to negotiate, empirical accounts of how this act comes into existence will become necessary. What does the Cirque du Soleil need to perform this controversial act? Which elements could be changed to readjust this fragile assemblage? In short, in order to propose compromises, in order to better compose, disputants will benefit from empirical accounts of the practices of trapeze;⁷ documenting what performers and entertainers cherish and fear and what they are attached to might allow constructive dissensions about the agency of what they produce to unfold.

尽管这个小小的想象例子有明显的局限性，但它表明，对可见性的要求与争议的增加有某种关联。当实践产品存在争议时，这些产品就不再被认为是充分的：因此，积极的不可见性可能会转变为消极的不可见性，而消极的不可见性本身需要实证解释——可以采取社会学调查的形式——争议可能会由此产生，谈判也可能会展开。当然，这些解释非常危险，因为它们本质上是以个人的名义发言的（Latour 2005，121-140）。为了使实践社区需要和珍视什么以及它们所依附的东西变得可见，可能为进一步的有争议的谈判建立共同基础的社会学解释需要克服许多考验：该解释是否使对实践者的工作至关重要的行为者变得可见？令人惊讶但有实证支持的联系是否展开？该解释是否为集体构成提出了新的把握？对这些问题中的任何一个回答“不”都会使社会学解释无法履行其最初的承诺。

Despite its obvious limits, this small imaginary example indicates that the request for visibility is somewhat correlated with the rise of controversies. When there are controversies over the products of practices, these products cannot be considered adequate anymore: positive invisibilities may thus switch to negative invisibilities that themselves call for empirical accounts—which can take the form of sociological investigations—on which disputes may arise and negotiations unfold. Of course, these accounts are very risky as they inherently speak in the name of individuals (Latour 2005, 121–140). To make visible what communities of practice need and cherish, and what they are attached to, the sociological account that may establish common grounds for further contentious negotiations would need to overcome many trials: Does the account make visible the actants that are crucial to the work of the practitioners? Do surprising but empirically supported connections unfold? Does the account propose new grips for collective composition? A single “no” to any one of these questions would make the sociological account fail to fulfill its initial commitment.

那么算法呢？不久前，这些设备几乎没有引起人们的注意。它们确实参与了权力关系的改变，但这些过程不是公共问题，或者只是在有限的程度上是公共问题。20 世纪 90 年代末，情况开始发生变化，社会学家开始质疑网络技术推动者提出的赋权和信息可访问性的论述。8例如^，霍夫曼和诺瓦克 (1998) 表明，美国网络技术的可访问性和使用在很大程度上取决于种族差异。劳伦斯和吉尔斯 (1999) 强调，与几乎无限访问的宣传口号相反，20 世纪 90 年代后期的搜索引擎只能索引网络上的一小部分。同样，Introna 和 Nissenbaum (2000) 强调了这些 20 世纪 90 年代后期搜索引擎用于对 URL 进行分类的启发式方法的暗中——并且可能有害——影响。随后的 9/11 事件后，人们开始批评程序和算法中的偏见——这一术语当时出现在批判性文献中^9——用于监视和预防性检测。例如，在研究数据挖掘技术的社会影响时，Gandy (2002) 警告说，它们是理性歧视的大门，可能会强化社会地位和群体成员之间的关联习惯。从政治经济角度来看，Zureik 和 Hindle (2004) 讨论了生物特征识别算法倾向于轻视社会分析、分类和民族群体排斥。另一个例子是 Introna 和 Wood (2004) 的工作：他们对面部识别算法的分析突出了这些设备的潜在偏见，当时这些设备通常被认为是公正的。 2010 年代初，这一条社会学研究路线引发了大量关于使用算法引起的歧视（例如 Kraemer、van Overveld 和 Peterson 2010；Gillepsie 2014 Steiner 2012）和隐形化（Bucher 2012；Bozdag 2013）的研究。

What about algorithms? Not so long ago, these devices attracted little attention. They were certainly involved in changing power relations, but these processes were not, or only to a limited extent, public issues. Things began to change in the late 1990s when sociologists started to question the discourse on empowerment and information accessibility put forward by the promoters of web technologies.⁸ Hoffman and Novak (1998) showed, for example, that the accessibility and use of web technologies in the United States were largely function of racial differences. Lawrence and Giles (1999) stressed that, contrary to the promotional rhetoric of almost unlimited access, the search engines available in the late 1990s were only able to index a small and oriented fraction of the web. In the same vein, Introna and Nissenbaum (2000) underlined the underground—and potentially harmful—influence of the heuristics used for the classification of URLs by these same late-1990s search engines. The post-9/11 period that followed focused on criticisms of biases in programs and algorithms—the term appeared at that time in the critical literature⁹—for surveillance and preventive detection. In his study of the social implications of data mining technologies, Gandy (2002) warned, for example, that they are the gateway to rational discrimination, potentially strengthening correlative habits between social status and group membership. From a political economy perspective, Zureik and Hindle (2004) discussed biometric algorithms’ propensity to trivialize social profiling, categorization, and exclusion of national groups. Another example is the work of Introna and Wood (2004): their analysis of facial recognition algorithms highlighted the potential biases of these devices, which were often, at that time, presented as impartial. This line of sociological research led, at the beginning of the 2010s, to numerous investigations on discriminations (e.g., Kraemer, van Overveld, and Peterson 2010; Gillepsie 2014 Steiner 2012) and invisibilizations (Bucher 2012; Bozdag 2013) induced by the use of algorithms.

近年来，这一研究方向一直在延续，越来越多的综合性研究揭示了算法对当代社会产生的对比性且往往值得怀疑的影响（例如，Crawford 和 Calo 2016；Noble 2018；O'Neil 2016；Pasquale 2015）。这些提高认识的努力也在媒体上报道，进一步使算法成为公众关注的问题（例如，Mazzotti 2017；Risen 和 Poitras 2017；Smith 2018）。这种动态——太复杂了，无法在本介绍¹⁰中彻底处理——导致了目前的局面，即集体世界不断受到算法争议的影响。在撰写本文时，快速浏览一下新闻就足以提醒我们这一点。英国警方即将使用一种新算法识别社交媒体上的网络仇恨犯罪 (Roberts 2017)？这很快引发了非营利组织“老大哥观察”的敌对反应，该组织准备“打击任何试图限制网络言论自由的行为” (Parker 2018)。学术期刊上发表了一种新算法，该算法可能可以根据人脸照片推断出人们的性取向 (Levin 2017)？男女同性恋反诽谤联盟很快谴责了这种“危险且有缺陷的研究，可能会对全世界的 LGBTQ 人群造成伤害” (Anderson 2017)。11^在儿子死产后，Facebook 的算法继续用育儿广告轰炸一位悲痛的女人 (Brockell 2018)？成千上万的推文很快谴责科技公司的性别歧视 (Mahdawi 2018)。每周都会出现一场关于新算法的后果（实际或潜在）的新争议，这些争议通常先于大数据、机器学习或最近的人工智能等定语名词的改变。

This research direction has continued in recent years, with increasingly comprehensive works revealing the contrasting, and often questionable, effects of algorithms on contemporary societies (e.g., Crawford and Calo 2016; Noble 2018; O’Neil 2016; Pasquale 2015). These awareness-raising efforts were also reported in the press, further making algorithms matters of public concern (e.g., Mazzotti 2017; Risen and Poitras 2017; Smith 2018). This dynamic—too complex to be thoroughly dealt with in this introduction¹⁰—has led to the current situation where the collective world is steadily affected by controversies over algorithms. A quick look at the news, at the time of writing, suffices to remind us of it. UK police is about to use a new algorithm to identify online hate crime on social media (Roberts 2017)? This soon triggers hostile reactions from the nonprofit organization “Big Brother Watch,” ready to “fight any attempt to curb free speech online” (Parker 2018). A new algorithm is published in an academic journal that can presumably deduce people’s sexuality from photographs of faces (Levin 2017)? The Gay & Lesbian Alliance Against Defamation soon condemns such a “dangerous and flawed research that could cause harm to LGBTQ people around the world” (Anderson 2017).¹¹ Facebook’s algorithm continues to bombard a grieved woman by parenting ads after the stillbirth of her son (Brockell 2018)? Thousands of tweets soon denounce gender bias from tech companies (Mahdawi 2018). Every week, a new dispute arises regarding the consequences—actual or potential—of new algorithms, often preceded by changing attributive nouns such as big data, machine learning, or more recently, artificial intelligence.

本书的预期相关性应根据当前有关算法代理的争议来考虑。继 Bechman 和 Bowker (2019)、Barocas 和 Selbst (2016) 以及 Grosman 和 Reigeluth (2019) 等作者的脚步（我将在本书后面回顾他们）之后，我的目标是提出一些智力工具来准备制定妥协方案。算法开发背后实践的不可见性确实不再被认为是积极的：由于它们是反复争论的对象，因此记录使它们得以存在的实际过程现在无疑很重要，或者至少很有趣。粗略地说，如果社会学已经成功地研究了算法的影响，那么现在是时候探究这些影响的原因了，无论这些影响多么分散和多样。需要填补一个空白；通过对计算机科学家和工程师如何培育算法的实证解释，我们或许可以为建设性的争论提供一些有风险但令人耳目一新的依据。12^算法设计者的需求、依恋和价值观——正如我有限的社会学解释所记录的那样——可能与其他需求、依恋和价值观相矛盾。但至少，在这些充满争议的日子里，争议各方可能会慢慢开始谈判，正如沃尔特·李普曼所说，“以自己的名义”（1982，91）。然而，在考虑如何有效地将这一探究应用于算法的实际形成之前，我需要进一步明确其政治层面。为此，我现在将快速绕道讨论我在这里用来限定我的冒险的非常规术语“宪法”。

The intended relevance of this book should be considered in the light of the current controversies over the agency of algorithms. Following in the footsteps of authors such as Bechman and Bowker (2019), Barocas and Selbst (2016), and Grosman and Reigeluth (2019)—to whom I shall return later in the book—my aim here is to propose intellectual tools to prepare the elaboration of compromises. The invisibility of the practices underlying the development of algorithms can indeed no longer be considered positive: as they are the object of repeated disputes, it is now certainly important, or at least interesting, to document the practical processes that enable them to come into existence. Roughly put, if sociology has looked, with a certain success, at the effects of algorithms, it is now time for it to inquire into the causes of these effects, however distributed and multiple they may be. A gap needs to be filled in; by means of empirical accounts of how computer scientists and engineers nurture algorithms, some risky yet refreshing grounds for constructive disputes may be provided.¹² The needs, attachments, and values of those who design algorithms—as documented by my limited sociological account—may contradict other needs, attachments, and values. But at least, in these days of controversies, parties in dispute may slowly start to negotiate, as Walter Lippmann says, “under their own colors” (1982, 91). Yet before considering how I intend to effectively run this inquiry into the practical formation of algorithms, I quickly need to further specify its political dimension. To do so, I shall now make a quick detour by discussing the unconventional term “constitution” I use here to qualify my venture.

为什么是“宪法”（而不是简单的“建设”）？

Why “Constitution” (And Not Simply “Construction”)?

在这篇导言的开头，我断言集体世界在不断地重新排列：异质实体永远不会停止相互关联，这些关联的融合暂时建立了新的事态。从这个（有争议的）本体论立场来看，世界并不是“在外面”，可以从某个外部角度来把握。相反，根据这种过程本体论，世界总是在变化；它是人类和非人类行为者之间关联的积极产物。

At the beginning of this introduction, I asserted that the collective world is constantly rearranged: heterogeneous entities never stop associating with each other, the blending of these associations temporarily establishing new states of affairs. From this (debatable) ontological position, it follows that the world is not “out there,” ready to be grasped from some outside standpoint. Instead, according to this processual ontology, the world is always becoming; it is the active product of associations between human and nonhuman actants.

然而，人们可以正确地指出，并非所有事物都是可以重新发明的。虽然有些关联会带来短暂的行为者（例如，欢呼、悲伤的眼泪、对某个笑话的笑声），但其他一些关联会带来更持久的行为者。许多填充/生成集体世界的实体都属于此类：马克·扎克伯格、火星、西岸监狱、Nutella 罐子——仅举几个我们在最初的小型 RT 中遇到的实体——都是相当持久的实体。由于这些行为者能够在其实例化之后继续存在，它们可能会反过来与其他行为者相关联，从而为集体世界的持续生成做出贡献。这些相对稳定的行为者具有一定的耐久性，使它们能够带来和引导正在发生的事情。

Yet one may rightly argue that everything is not always reinvented. While some associations bring about ephemeral actants (e.g., a cry of joy, tears of sadness, laughs at some joke), some other associations bring about actants that are more enduring. Many entities that populate/generate the collective world are of this sort: Mark Zuckerberg, the planet Mars, West Bank jails, Nutella jars—just to mention some entities we encountered in our small initial RTs—are quite enduring entities. Such actants, thanks to their ability to live on beyond the here and now of their instantiation, may in turn associate themselves with other actants, thus contributing to the continuous generation of the collective world. Such relatively stable actants possess some durability that allows them to bring about and orient what is becoming.

如果我们继续考虑行为者之间的差异，我们很快就会注意到，一些持久的行为者可以或多或少轻松地从一个地方移动到另一个地方。让我们继续用熟悉的实体来说明这一点。如果我们考虑火星和西岸监狱，这些实体看起来相当静态。它们很难与能够使它们偏离初始轨迹的行为者联系起来：如果没有重要的动员努力，火星和西岸监狱将停留在原地。但马克·扎克伯格的情况并非如此，他一旦与“鞋子”、“汽车”或“道路”等行为者联系起来，就可以显著改变他的初始轨迹，进而将自己与最初与他疏远的其他行为者联系起来。然而，很大程度上由于他的身体外壳，马克·扎克伯格的相对机动性是相当昂贵的：为了让他以某种方式继续做马克·扎克伯格，为了让他在移动时保持大部分耐久性，他需要与许多其他行动者（例如，氧气、食物、腿部空间、咖啡休息）相关联，以保护他免受太大的改变。在 Nutella 罐子的情况下，故事有点不同。它们也需要与其他行为者关联，才能偏离其初始轨迹（例如，供应链经理、铁路线、销售合同、送货员）。但与马克·扎克伯格相反，我们可以合理地假设 Nutella 罐子的改变较慢：由于它们适当的物质性，由于它们自己的媒介，它们可以被储存、堆积和处理而不会发生显着改变。在我们示例性的耐用实体中，Nutella 罐子似乎是最耐用和最机动的：与火星、西岸监狱甚至马克·扎克伯格相比——并且当提供足够的关联时——这些罐子可以从一个地方移动到另一个地方而不会发生太大改变。

If we continue considering differences among actants, we quickly notice that some durable actants can move from one place to another more or less easily. Let us keep on using familiar entities to illustrate this point. If we consider the planet Mars and West Bank jails, these entities appear rather static. It is difficult for them to associate with actants capable of making them deviate from their initial trajectories: without important mobilization efforts, the planet Mars and West Bank jails will just stay where they are. This is not quite the case for Mark Zuckerberg who, once associated with actants such as “shoes,” “cars,” or “roads,” can markedly change his initial trajectory and, in turn, associate himself with other actants that were at first distant from him. Yet, largely due to his body envelope, Mark Zuckerberg’s relative mobility is rather costly: in order for him to somehow keep on being Mark Zuckerberg, in order for him to maintain most of his durability while he is moving, he would need to associate with many other actants (e.g., oxygen, food, space for his legs, coffee breaks) protecting him from being too much altered. In the case of Nutella jars, the story is a bit different. They too need to associate with other actants to deviate from their initial trajectories (e.g., supply chain managers, railway lines, sale contracts, delivery people). But contrary to Mark Zuckerberg, one can make the fair assumption that Nutella jars’ alteration is slower: due to their proper materiality, due to their own medium, they can, for example, be stored, piled up, and handled without being significantly transformed. Among our exemplary durable entities, Nutella jars seem then the most durable and mobile: when compared to the planet Mars, West Bank jails, or even Mark Zuckerberg—and when provided adequate associations—these jars can move from one place to another without being too much altered.

持久性和流动性累积起来，就是不平凡的特征：兼具这两种能力的实体更有可能与其他实体相关联，从而积极地为集体世界的生成做出贡献。但有一类非常特殊的实体还累积了另一种能力，这使得它们无疑是所有实体中最具世界生成能力的。这些实体有不同的名称：杰克·古迪 (Jack Goody) 称它们为“图形对象”(1977)；布鲁诺·拉图尔和史蒂夫·伍尔加 (Steve Woolgar) 称它们为“铭文”(1986, 43–91)；多萝西·史密斯 (Dorothy Smith) 称它们为“账户”或“文件”(1974)。但无论如何称呼，社会学家长期以来一直强调这些行为者的迷人能力，即持久性和流动性，以及带有其他行为者的一些特征——或行为者之间的其他关联。这本质上就是文本、表格、图表或图画的作用：由于特定习惯、规则和技术的存在和不断维护——杰罗姆·丹尼斯（ Jérôme Denis，2018）称之为“脚本基础设施”，这些通常持久且可移动的铭文可以承载某些行为因素和联想，并在其他地方再次呈现（重新呈现）。这种（部分）行为因素的脚本传输——本身需要许多其他行为因素的展开——反过来可能会在已经发生的事情和即将发生的事情之间建立联系。这听起来像是一个奇怪的说法，但这种现象实际上非常普遍：每次我读到《纽约时报》的文章，过去发生的事情（一些事件）和现在正在发生的事情（我，考虑这个事件，并最终对其做出反应）之间就会建立联系。当然，这种联系、这种链接已被格式化，以便承载在我正在考虑的铭文（这里是报纸文章）的特定物质性中。因此，这种联系始终是部分但可能忠实的、知情的版本。当我阅读《纽约时报》时，我并没有看到移民努力争取欧洲的恶劣条件；我看到一个平面，上面写着那些移民的文字；这种再现在我心中引发了无助、羞耻和绝望的感觉，这些转瞬即逝的行为者反过来又会为集体世界的不断生成做出贡献（尽管微不足道）。为了限定铭文承载行为者-联想的某些属性的能力，并在时间和地点之间建立格式化但生成的联系，我将使用“可再现性”一词。铭文不仅是持久和可移动的行为者，因此也是可再现的：它们可以——与合适的基础设施一起——承载、传输和展示不仅仅是它们自己的属性。

When cumulated, durability and mobility are nontrivial characteristics: entities that combine both abilities are more likely to associate with other entities, thus actively contributing to the generation of the collective world. But a very special category of entities cumulates another ability that makes them certainly the most world-generative of all. These entities go by different names: Jack Goody calls them “graphical objects” (1977); Bruno Latour and Steve Woolgar call them “inscriptions” (1986, 43–91); Dorothy Smith calls them “accounts” or “documents” (1974). But no matter how these are labeled, sociologists have long emphasized on these actants’ fascinating capacity to be durable and mobile and to carry with them some characteristics of other actants—or of other associations between actants. This is essentially what texts, tables, graphs, or drawings do: thanks to the presence and constant maintenance of specific habits, rules, and technologies—what Jérôme Denis (2018) calls scriptural infrastructures—these often durable and mobile inscriptions can host some aspects of actants and associations and present them again (re-present) somewhere else. This scriptural transport of (part of) actants—that itself necessitates many other actants to unfold—may in turn create a link between what has happened and what is to become. This sounds like an odd statement, but such a phenomenon is in fact very common: Every time I read a New York Times article, a connection is made between what has happened in the past (some events) and what is happening now (me, considering this event and, eventually, reacting to it). Of course, this connection, this link has been formatted in order to be hosted in the specific materiality of the inscription I am considering (here, the newspaper article). Such a link is thus always a partial, but potentially faithful, in-formed version of what has happened. When I’m reading the New York Times, I don’t see migrants struggling to reach Europe in horrendous conditions; I see a flat surface with words that re-present me those migrants; this re-presentation triggering in me feelings of helplessness, shame, and despair, evanescent actants that will, in turn, contribute to the continuous generation of the collective world (though quite insignificantly). To qualify inscriptions’ capacity to carry some properties of actants-associations and establish formatted yet generative connections between times and locations, I shall use the term “re-presentability.” More than just being durable and mobile actants, inscriptions are thus also re-presentable: they can—together with suitable infrastructures—carry, transport, and display properties that are not only theirs.

耐久性、移动性、可再现性：这些能力不容小觑。尽管铭文通常看起来不起眼（数字列表、图画、文章、表格、图表），但它们对我们世界的塑造却发挥着重要作用。一种新的分子的出现彻底改变了我们对人类下丘脑的理解？正如拉图尔和伍尔加（1986）所记录的那样，这种易于联想的行为在很大程度上源于实验室内部和实验室之间收集、积累、汇编和比较的铭文。一种新的管理技术开始将企业活动与单一的任意标准保持一致？正如泰维诺（1984）和耶茨（1989）所提出的，这种泰勒主义规范化及其后果严重依赖于测量、编码和公平方法，这些方法的圣经流通允许对工人进行集中控制。一种新算法的发布可能会点燃数字图像处理研究的原始途径？正如我将在整本书中试图展示的那样，这种行为因素的形成在很大程度上归功于许多不同类型的铭文的产生、传播、转化和编纂。我们将在适当的时候更彻底地研究铭文的世界生成能力（特别是在第 4、5 和 6 章中）。现在，可以说这些持久、移动和可再现的行为因素对不断发生的事情做出了很大贡献。

Durability, mobility, re-presentability: these are capacities not to be underestimated. Inscriptions, despite their often-modest appearances (lists of numbers, drawings, articles, tables, graphs), greatly participate in the shaping of our world. A new molecule appears that revolutionizes our understanding of the human hypothalamus? As well documented by Latour and Woolgar (1986), such an association-prone actant derives, to a large extent, from inscriptions assembled, accumulated, compiled, and compared within and between laboratories. A new management technique starts to align corporate activities to a single arbitrary standard? As proposed by Thévenot (1984) and Yates (1989), such Taylorist normalization—and its consequences—heavily relies on measures, coding, and equity methods whose scriptural circulation allows the centralization of control over the workers. A new algorithm is published that may ignite original avenues of research in digital image processing? As I will try to show throughout this book, the formation of such an actant owes a great deal to the production, circulation, transformation, and compilation of many different types of inscriptions. We will more thoroughly examine the world-generative capacity of inscriptions in due time (especially in chapters 4, 5, and 6). For now, suffice it to say that these durable, mobile, and re-presentable actants contribute a lot to what is constantly happening.

但无论其生成能力如何，“铭文”都不是独立存在的：它们显然需要在开始流传之前被制作出来。从这个意义上说，每个铭文都需要被铭刻。提取联想（或“事件”；在这一点上，这两个术语是等同的）的某些方面并将它们重新呈现在平面、耐用的媒介上根本不是显而易见的：事件的哪一部分应该保留和写下来？应该使用什么语言？应该遵循什么协议来比较这个铭文与其他一些人合作，并反过来产生新的编纂铭文？考虑到铭文的世界生成潜力，这些都是主要问题，大多数时候都得到组织和专业实践的支持，这些实践有自己的目标、规则和原则，每天都有数亿人和工具参与其中。这种以制作铭文为导向的工作，最终利用其世界生成潜力，这就是多萝西·史密斯 (1974) 所说的“记录现实的结构”。13^这种结构具有高度的政治性。

But whatever their generative power, “inscriptions” do not exist by themselves: they obviously need to be produced before they start to circulate. In that sense, every inscription needs to be inscribed. Extracting some aspects of associations (or “events”; at this point, both terms are equivalent) and re-presenting them on flat, durable media is not at all evident: What part of the event shall be kept and written down? What language shall be used? What protocol shall be followed to later compare this inscription with some others and produce, in turn, new compiled inscriptions? Considering the world-generative potential of inscriptions, these are major issues, most of time supported by organizational and professional practices with their own goals, rules, and principles that every day engage hundreds of millions of people and instruments. This oriented work consisting in producing inscriptions and, eventually, capitalizing on their world-generative potential is what Dorothy Smith (1974) calls “the fabric of documentary reality.”¹³ And this fabric is highly political.

为了说明她的观点，史密斯举了一个先验的平凡例子——出生证明。事实上，在报告上记录出生既不明显也不中立。它是一种组织和专业实践的产物，这种实践以非常特殊的方式塑造了出生及其记录，与母亲和父亲可能想要记住的方式截然不同。正如她所说：

To illustrate her point, Smith takes the a priori mundane example of birth certificates. Inscribing a birth on a report is, in fact, not evident nor neutral. It is the product of an organizational and professional practice that shapes births and their accounts in very peculiar terms, very different from, say, how mothers and fathers may want to remember it. As she put it:

“杰西·弗兰克出生于 1963 年 7 月 9 日”在这方面似乎是最明确的。但是，当我们研究它是如何被伪造时，就会发现它仅仅是一份记录，这是它被伪造的一部分。母亲和父亲可能希望记住杰西·弗兰克的出生对他们意味着什么，但这些信息都存储在其他地方，并且在记录机构的实践中被明确丢弃，因为它们无关紧要。后者只关心在特定个人的出生（一个实际事件）与姓名和确定该个人位置所必需的某些社会坐标（她父母的名字、她的出生地等）之间建立经过认证的永久联系。（Smith 1974，264）

“Jessie Franck was born on July 9th, 1963” appears maximally unequivocal in this respect. But as we examine how it has been fabricated it becomes apparent that its character as merely a record is part of how it has been contrived. Everything that a mother and a father might want to have remembered as how the birth of Jessie Franck was for them is stored elsewhere and is specifically discarded as irrelevant in the practices of the recording agency. The latter is concerned only to set up a certified and permanent link between the birth of a particular individual—an actual event, and a name and certain social coordinates essential to locating that individual—the names of her parents, where she was born, etc. (Smith 1974, 264)

出生证明非常有选择性——它们只保留出生事件的一小部分——这种选择面向这种简明铭文的潜力——它们的特征反过来可以用于身份识别目的或政府统计。此外，作为可以在其他空间重新动员的铭文，出生证明及其预期用途会形成特定的出生版本，在许多情况下，这些版本会强加于其他并发版本。尽管这些流传的铭文的起源非常片面和党派性，但它们将成为其他铭文的支点，逐步建立正式、事实和所谓的“中立”出生版本。

Birth certificates are very selective—they only keep a very small part of birth events—and this selection is oriented toward the potential of such concise inscriptions—their features can, in turn, be used for identification purposes or government statistics. Moreover, as being inscriptions that can be remobilized in other spaces, birth certificates and their desired purposes make a specific version of births that will, in many cases, impose on other concurrent versions. Despite their very partial and partisan origins, these circulating inscriptions will form a fulcrum for other inscriptions, progressively establishing formal, factual, and so-called “neutral” versions of births.

铭文实践的政治方面旨在对事件做出部分党派化解释，这不仅与行政有关。史密斯论点的力量在于，它也适用于任何铭文，因为从物质上讲，不可能完全铭刻一个事件的所有细节：需要选择保留（和格式化）什么以及忽略什么。铭文作为世界创造者所获得的东西，也作为世界叛徒所失去的东西，后者甚至是前者的条件^。14

This political aspect of inscription practices which aim to make partial partisan versions of events does not only concern administration. The power of Smith’s argument lies in that it is also applicable to any inscription as it is materially impossible to fully inscribe an event in all its subtleties: choices need to be made regarding what will be kept (and formatted) and what will be ignored. What inscriptions gain as world-generators also lose as world-betrayers, the latter being even a condition to the former.¹⁴

考虑到这些因素，现在让我们回到本书。我不是说过这本书旨在成为一部社会学著作吗？我不是说过这本书旨在解释逐渐形成我们称之为算法的设备的关联吗？在这一点上，可以进一步明确这些断言。社会学作为一项专业活动，其内容是制作关于关联（socius ）的专业文本（ logos），但它也逃脱不了我现在称之为“多萝西·史密斯定律”的束缚：无论它多么具有描述性，社会学都会通过铭文创造部分现实，而损害其他现实。对管理者（Desrosières 2010）、经济学家（MacKenzie、Muniesa 和 Siu 2007）或科学家（Latour 1987）来说是正确的，对社会学家来说也是如此：他们在通过文本描述现实的同时，也在实践这些现实。

With these elements in mind, let us now come back to this present book. Have I not said it intends to be a sociological work? Have I not said it intends to account for associations that progressively form devices we call algorithms? At this point, these assertions can be further specified. Sociology, as a professional activity that consists in producing specialized texts (logos) about associations (socius), does not escape what I shall now call “Dorothy Smith’s law”: however descriptive it is, sociology brings into being—by means of inscriptions—partial realities to the detriment of other realities. What is true for administrators (Desrosières 2010), economists (MacKenzie, Muniesa, and Siu 2007), or scientists (Latour 1987) is also true for sociologists: while describing realities by means of texts, they also enact these realities.

正如 Law 和 Urry (2004, 396) 很好地总结的那样，没有无辜：¹⁵一个文本，无论多么忠实——有些文本肯定比其他文本更忠实——也是一厢情愿的。然后我必须承认，我在这本书中想要做的不仅仅是描述在特定的、与算法相关的情形中发生的事情：由于这本书本身就是一个文本铭文，它也是一种试图制定一个损害其他制定世界的世界的尝试。因此，我的举动是分析性的和政治性的：它旨在对算法如何出现做出描述性说明——我们可以保留这一点——但同时也提出一个关于算法现实的新版本。上一节介绍了这种分析政治举措背后的动机：在围绕算法代理权存在争议的今天，对其内部组成部分进行精炼的——但格式化并因此在本质上受到限制的——描述可以为关于算法和与算法进行建设性的争论奠定基础。

As Law and Urry (2004, 396) well summarized it, there is no innocence:¹⁵ a text, however faithful—and some texts are definitely more faithful than others—is also a wishful accomplishment. I must then admit that what I intend to do in this book is not only describing what happens in particular, algorithm-related, situations: due to this book’s very existence as a textual inscription, it is also an attempt at enacting a world to the detriment of other enacted worlds. My gesture is thus analytical and political: it aims to produce a descriptive account of how algorithms come into existence—we can keep that—but also, and in the same movement, to propose a new version of their realities. The motivation behind this analytico-political move were presented in the previous section: in these days of controversies over the agency of algorithms, a refined—yet formatted and thus intrinsically limited—account of their inner components may establish grounds for constructive disputes about and with algorithms.

回到本节的标题，我认为“建构”这一经典概念不能很好地表达这种冒险。建构无疑是社会学的一个有用术语，因为它为许多自然化问题提供了有价值的批评：关于性别建构（Lorber and Farrell 1991）、父权制（Lerner 1986）或母性（Badinter 1981）的研究，仅举几例经典之作，都具有极大的解放性。但考虑到 STS 和一般社会学的最新发展，建构似乎存在两面性：虽然它很好地表达了其描述性愿望——展示了结果是如何产生的——但它也倾向于隐藏其政治主张——产生损害他人的现实。16^由于它倾向于将“多萝西·史密斯定律”隐藏在为了掩盖分析野心，我认为放弃使用“建构”一词来限定我的整体姿态是更明智的做法。

To come back to the title of this section, I assume the classical notion of “construction” does not well express such a venture. Construction has been for sure a useful term for sociology as it has equipped many valuable critiques of naturalized matters: studies on the construction of gender (Lorber and Farrell 1991), patriarchy (Lerner 1986), or maternity (Badinter 1981), just to mention some classics, have all been wonderfully liberating. But considering recent developments in STS and sociology in general, it appears that construction suffers from being two-faced: while it well expresses its descriptive aspirations—showing how results have been produced—it also tends to hide its political claims—generating realities to the detriment of others.¹⁶ Due to its propensity to hide “Dorothy Smith’s law” under the cover of analytical ambitions, I consider it wiser to renounce using the term “construction” to qualify my overall gesture.

我并不是第一个否定建构的社会学家。事实上，这是一个相当流行的举动，其动机与上述论点大致相同。劳和厄里（2004）更喜欢使用“制定”，因为它更好地表达了描述性冒险的表现性。拉图尔（2013）受到苏里奥（[1943] 2015）的启发，诉诸“建立”，因为它强调了实际的、连续的集合的脆弱性。英戈尔德（2014）在罗蒂（1980）之后，优先考虑“教化”，因为它强调即将发生的事情的连续性和从未完全实现的方面。所有这些概念肯定都是建构的有趣替代品。但冒着引入已经很丰富的社会学术语的风险，我选择在这里使用“宪法”的概念，因为它具有显著的优势，即本身就包含双重含义：既是某事发生的过程，也是一份倡导权利和特权的文件。这里存在着一种有趣的张力，可能会让人想起我的姿态的假定矛盾：描述和争论。此外，由于宪法永远不会一劳永逸地固定下来（它可以被修改、完成、废除），这个概念迫使我们认识到我的冒险必然是不完整的，我试图在这里实现的三项活动——实地调查、编程和制定（显然，稍后会详细介绍）——必须被视为部分和暂时的。只要有经验材料的支持，就可以将更多的动名词冠词添加到算法的现有组成行为中。

I am not the first sociologist to dismiss construction. It is in fact quite a popular move, motivated by more or less the same arguments as presented above. Law and Urry (2004) prefer to use “enactment” as it better expresses the performativity of descriptive ventures. Latour (2013), inspired by Souriau ([1943] 2015), has recourse to “instauration” as it underlines the fragility of practical, succeeding assemblages. Ingold (2014), in the wake of Rorty (1980), gives priority to “edification” as it stresses the continuous and never fully achieved aspect of what is about to happen. All these notions are surely interesting alternatives to construction. But at the risk of feeding in a sociological jargon already well supplied, I choose here to use the notion of “constitution” as it has the significant advantage of containing natively a double signification: a process by which something occurs as well as a document advocating for rights and prerogatives. Here lies an interesting tension that may recall the assumed ambivalence of my gesture: describing and contesting. Moreover, as a constitution is never fixed once and for all (it can be amended, completed, abolished), the notion forces us to recognize the necessary incompleteness of my venture, the three activities that I try to put into existence here—ground-truthing, programming, and formulating (more on this later, obviously)—must be considered partial and temporary. Many more gerund articles, as long as they are supported by empirical materials, can be potentially added to the present constituent act of algorithms.

出于所有这些原因，本书的标题《算法的构成》应被理解为将算法付诸文字并使其存在——既是经验性的又是行动性的。在探究的最后，鉴于所考虑的因素，我将在借用安东尼奥·内格里（Antonio Negri）关于“制宪权”的著作（1999）的部分中回顾这种分析/叛乱姿态的含义。现在，让我们通过使用“构成”一词来注意并接受这种矛盾心理，这不断提醒我们这一探究的两极性。

For all these reasons, this book’s title The Constitution of Algorithms should be understood as the putting into text and existence—simultaneously empirical and activist—of what algorithms shall be. At the very end of the inquiry, in light of the accounted elements, I will come back to the implications of this analytical/insurrectional gesture in a section borrowing from Antonio Negri’s (1999) work on “constituent power.” For now, let us just note and accept this ambivalence by using the term constitution, a constant reminder of this inquiry’s bipolarity.

实验室研究

A Laboratory Study

在这一点上，我别无选择，只能请读者跟随我——至少是暂时的——假设在关于算法的代理，设计、塑造和传播算法所需工作的不可见性是负面的，因为它阻止了争议方有共同的谈判基础。我们还假设，提出这种基础的一种方式，从而提出建设性的争议和创作尝试，可能是进行社会学调查，以使算法存在所需的工作实践可见。最后，让我们假设这本书是对这样一种调查的尝试，作为世界生成的铭文，它不能不成为算法的部分、党派和开放式（同时也是忠实和经验的）的构成。如果我们接受这些有争议的假设，下一个问题可能是：我如何有效地进行这种部分的、经验的和积极的调查？我可以用什么材料作为它的基础？

At this point, I have no other choice than to ask the reader to follow me—at least temporarily—in assuming that in these days of controversies over the agency of algorithms, the invisibility of the work required to design, shape, and diffuse them is negative as it prevents disputing parties from having common grounds for negotiations. Let us also assume that one way to propose such grounds, and thus to suggest constructive disputes and composition attempts, could be to conduct sociological inquiries in order to make visible the work practices required to make algorithms come into existence. Let us finally assume that this volume is an attempt at such an inquiry that, in its capacity as a world-generative inscription, cannot but be a partial, partisan, and open-ended (while also faithful and empirical) constitution of algorithms. If we accept these debatable assumptions, the next question could be: How can I effectively run such a partial, empirical, and activist inquiry? On what materials can I ground it?

使用现成的资料来源，例如描述算法内部工作原理的许多学术论文和手册，会很诱人。事实上，一些 STS 学者在一些非常有趣的作品中就是这么做的。17^然而，我有理由相信，仅仅使用这些资料会暗中助长算法组件的负面隐形。关于发表在学术期刊上的计算机科学论文，说这些文献是错误的当然是不正确的：相反，它证明了即将成为科学真理的东西。18但正如许多重要的科学研究表明的^那样，这些科学出版物往往报告过程的结果，而不是导致这些结果的实际活动。在这种情况下，仅仅使用学术出版物来使算法的形成可见是有问题的，因为这些文档本身是由未说明的元素支持和框架的。迈克尔·林奇 (Michael Lynch) (1985) 很好地总结了科学出版物分析中固有的这个问题：

It would be tempting to use readily available sources, such as the many academic papers and manuals describing the internal workings of algorithms. This is in fact what several STS scholars have done in some very interesting works.¹⁷ However, I have reasons to believe that the sole use of these sources surreptitiously contributes to the perpetuation of the negative invisibility of algorithms’ components. Regarding computer science papers published in academic journals, it would, of course, be incorrect to say that this literature is erroneous: on the contrary, it attests to what is about to, perhaps, become scientifically true.¹⁸ But as many important science studies have shown, these scientific publications tend to report the results of processes, not the practical activities that led to those results. Under these conditions, it is problematic to solely use academic publications to make the formation of algorithms visible since these documents are themselves supported and framed by unstated elements. Michael Lynch (1985) well summarized this problem inherent in the analysis of scientific publications:

[科学研究论文的方法部分] 为已经熟练的从业者提供了循序渐进的行为准则，让他们能够吸收常识和未制定但具体为科学的探究实践的无限组合。当科学研究依赖实验室探究的文献残余作为科学工作的可观察和可分析的存在时，这些未制定的实践必然会被排除在研究领域之外。（Lynch 1985，3）

[Methods sections of scientific research papers] supply step-by-step maxims of conduct for the already competent practitioner to assimilate within an indefinite mix of common sense and unformulated, but specifically scientific, practices of inquiry. These unformulated practices are necessarily omitted from the domain of study when science studies rely upon the literary residues of laboratory inquiry as the observable and analyzable presence of scientific work. (Lynch 1985, 3)

此外，由于我们将在本书中介绍的复杂原因，学术论文的作者也倾向于为他们的算法辩护，反对并发算法。在科学期刊上发表的主张确实是针对其他主张的，旨在获得读者的支持。因此标题技术的重要性在于“将文本布局得让读者无论身在何处都只有一条路可走”（Latour 1987，57）。这些信念习惯及其提供的额外必要性——建立客观构造的基本要素——倾向于净化算法的科学描述，而这些描述中有许多不同的元素促成了文本的存在。因此，当依靠这些文档来分析计算机计算方法时，犹豫、怀疑和“非同寻常”的设备和写作往往会逃脱分析师的视线^。19

Moreover, for entangled reasons we will cover throughout this book, authors of academic papers tend also to defend their algorithms against concurrent algorithms. A claim published in a scientific journal is indeed directed against other claims and is intended to obtain the reader’s support. Hence the importance of captation techniques that aim “to lay out the text so that wherever the reader is there is only one way to go” (Latour 1987, 57). These conviction habits and the additional necessity they provide—essential elements to establish objective constructions—tend to purify the scientific accounts of algorithms of the many disparate elements that have contributed to their textual existence. When relying on these documents to analyze computerized methods of calculation, it is therefore the hesitations, doubts, and “infra-ordinary” equipment and writings that tend to escape the analyst’s gaze.¹⁹

但是，那些教我们如何设计算法的大量手册呢？²⁰它们不是提供了如何组装计算机计算方法的描述吗？从这个意义上说，它们不是算法与它们所塑造的集体世界之间的连接点吗？这些教学资源对于向学生和新手灌输计算机计算方法的基本组成部分当然至关重要，而这些组成部分对于他们的社会学分析至关重要。然而，正如露西·萨奇曼（Lucy Suchman，1995）提醒我们的那样，这些资源从定义上讲，是关于应该如何完成工作的规范性说明，而不是关于如何有效地完成工作的规范性说明。这是一个至关重要但经常被遗忘的精确性：“[这些]规范性说明代表了理想化和典型化。因此，它们的写作依赖于偶然性和差异性的删除”（Suchman，1995，61）。手册不是解释在平凡情况下正在做什么，而是解释应该做什么。它们是（重要的）强制性处方，而不是基于经验的实践说明。²¹我认为，这是当代研究的主要局限性，这些研究主要依赖算法设计教科书和课程：它们告知当代教育家希望算法如何构建，而不是告知这些算法在日常生活中如何构建。这些研究虽然非常有趣，但并没有通过解释他们的工作来接近计算机科学家，反而倾向于将他们拉得更远^。22

But what about the numerous manuals that teach us how to design algorithms?²⁰ Do they not provide descriptions of how to assemble computerized methods of calculation? Are they not, in that sense, connectors between algorithms and the collective world they contribute to shaping? These pedagogical resources are certainly crucial to inculcate students and newcomers with the basic components of computerized methods of calculation, which are essential to their sociological analysis. Yet, as Lucy Suchman (1995) reminded us, these resources are, by definition, normative accounts of how work should be done, not of how work is effectively done. This is a crucial but often forgotten precision: “[These] normative accounts represent idealization and typifications. As such, they depend for their writing on the deletion of contingencies and differences” (Suchman 1995, 61). Instead of accounting for what it is being done during mundane situations, manuals account for what ought to be done. They are (important) peremptory recipes, not empirically grounded accounts of practices.²¹ This is, I believe, the main limitation of contemporary studies that rely mainly upon textbooks and classes on algorithmic design: they inform about how contemporary pedagogues want algorithms to be constructed, not on how these algorithms are constructed on a day-to-day basis. Instead of getting closer to computer scientists by accounting for their work, these studies, otherwise very interesting, tend to move them further away.²²

因此，学术论文和手册是需要谨慎处理的资料来源。但是，如何了解这些仍然有用和重要的资料对隐藏起来有什么作用？如何获得更高清晰度但本质上仍然有限的算法所需工作的图景？幸运的是，对于这个非常具体的目的，我可以依靠一种经过验证的 STS 分析类型，通常被称为“实验室研究”。第一批此类研究出现在 20 世纪 70 年代，主要是在美国。在从某种意义上说，当时的集体（西方）世界与我们今天所经历的世界并无太大不同：关于机构类型的争议不断出现。但这些争议主要涉及的不是算法，而是生命科学、物理学和神经学中经常出现的科学事实。由于许多原因（这些原因太过复杂，无法在本导言中讨论）²³，几位学者认为有必要通过社会学解释自然科学家试图制造认证知识的平凡做法来消除科学事实的欺骗性（Collins 1975；Knorr-Cetina 1981；Lynch 1985；Latour and Woolgar 1986）。这些学者的方法相当激进：为了回应认识论的权威戒律，这些作者借用了民族志的现场分析视角来记录“科学的软肋”（Edge 1976）。正如 Latour 和 Woolgar 所说：

Academic papers and manuals are therefore sources that should be handled with precautions. But how to reach what these sources, which remain useful and important, contribute to keeping out of sight? How to get a higher definition, yet still intrinsically limited, picture of the work required to assemble algorithms? Fortunately, for this very specific purpose, I can rely on a proven STS analytical genre often labeled “laboratory study.” The first such studies appeared in the 1970s, mostly in the United States. In a sense, the collective (Western) world was at that time not so dissimilar to the one we are experiencing today: controversies about types of agencies were arising continuously. But instead of algorithms, these controversies mostly concerned scientific facts often developed in life science, physics, and neurology. For many reasons that are too entangled to be discussed in this introduction,²³ several scholars felt the need to deflate the delusive aspect of scientific facts by sociologically accounting for mundane practices of natural scientists trying to manufacture certified knowledge (Collins 1975; Knorr-Cetina 1981; Lynch 1985; Latour and Woolgar 1986). The method of these scholars was quite radical: in reaction to the authoritative precepts of epistemology, these authors borrowed from ethnography its in situ analytical perspective to document “the soft underbelly of science” (Edge 1976). As Latour and Woolgar put it:

我们设想了一种类似于象牙海岸的勇敢探险家的研究程序，他通过与部落成员一起生活、分享他们的艰辛并几乎成为他们中的一员，研究了“野蛮人”的信仰体系或物质生产，最终带回了一系列观察结果，可以作为初步研究报告提交。……我们特别重视收集和描述在特定环境中获得的科学活动观察结果。（1986，28；重点为原文所加）

We envisaged a research procedure analogous with that of an intrepid explorer of the Ivory Coast, who, having studied the belief system or material production of “savage minds” by living with tribesmen, sharing their hardship and almost becoming one of them, eventually returns with a body of observations which he can present as a preliminary research report. … We attach particular importance to the collection and description of observations of scientific activity obtained in a particular setting. (1986, 28; emphasis in the original)

这些实验室民族志学者——他们积极参与发起《科学与技术研究》——决定从日常行为和工作实践入手，记录并展示科学事实是如何逐步形成的，而不是从科学理论、思想或“理性法则”入手。在 20 世纪 80 年代开创性的实验室研究之后，出现了几本关于物理学家（Traweek 1992；Sormani 2014）和设计工程师（Vinck 2003）实践的专著，每次都提供了富有洞察力的新结果。我们将在适当的时候介绍其中一些结果。现在，可以说，当前的社会学研究几乎完全基于这些著作。但这具体意味着什么呢？

Instead of starting from scientific theories, minds, or “laws of Reason,” these laboratory ethnographers—who actively participated in the launching of Science and Technology Studies—decided to start from mundane actions and work practices to document and make visible how scientific facts were progressively assembled. Several other monographs accounting for the practices of physicists (Traweek 1992; Sormani 2014) and design engineers (Vinck 2003) followed the seminal 1980s laboratory studies, each time providing insightful new results. We will cover some of these results in due time. For now, suffice it to say that the present sociological inquiry is based almost entirely on these works. But what does that concretely imply?

首先，这意味着要找到人们每天工作以组装算法的地方。就我而言，这种定位工作并不困难，因为我在机构上靠近欧洲技术学院，该学院有大约 20 个计算机科学实验室，每天都在努力提出新算法，并使其在更广泛的学术和工业网络中传播。一项更艰巨的任务是说服其中一个实验室的主任让我将算法的实际形成描述为一位“勇敢的探险家”。幸运的是，与建立新的数字人文研究所有关的机构运动使我能够与一位对跨学科持开放态度的计算机科学教授分享我的研究抱负。24^经过几次尝试，我得以成为她数字图像处理实验室的一员，为期两年半，从 2013 年 11 月到 2016 年 3 月。这不是被动的时刻：根据实验室研究的分析类型以及我作为正式成员所属的实验室的规则，我必须参与实验室的生活，从而变得有些能干。虽然我逐渐掌握的技能当然不会让我成为一名计算机科学家，但它们对于充分谈论我的新同事关心的问题仍然至关重要。但参与和讨论是不够的：我还必须写下、收集和汇编我所做、所见和所讨论的内容。具体来说，这意味着做大量笔记。讨论、会议、演讲、行动：我所经历的一切，理想情况下，都必须记录下来，引用笔记本和计算机文档，以便日后检索、比较、取样和分析。这项全职数据汇编工作意味着最后一步：在我留在计算机科学实验室之后——在此期间，我参与了项目，与同事进行了讨论，观察了他们的工作，尽可能多地记录下来，并就我的初步结果进行了演讲（这些过程深深地改变了我和我现在从事的社会学研究）——我必须回到自己的研究社区，更深入地研究收集到的材料，并撰写一份调查报告，这份报告逐渐成为了现在这本书。

It first implies locating places where individuals work daily to assemble algorithms. For my case, this localization exercise was not very difficult as I was institutionally close to a European technical institute with about twenty computer science laboratories working every day to propose new algorithms and to make them circulate in broader academic and industrial networks. A more arduous task was to convince the director of one these laboratories to let me describe the practical shaping of algorithms as an “intrepid explorer.” Fortunately, institutional movements related to the establishment of a new institute of digital humanities enabled me to share my research ambitions with a computer science professor open to interdisciplinarity.²⁴ And after several trials, I could be part of her laboratory of digital image processing for two and half years, from November 2013 to March 2016. These were no passive moments: as required by the analytical genre of laboratory studies and also by the rules of the laboratory to which I was affiliated as full member, I had to participate in the life of the laboratory and thus become somewhat competent. Although the skills I progressively acquired certainly did not make me become a computer scientist, they were nonetheless crucial for speaking adequately about issues that mattered to my new colleagues. But participating and discussing were not enough: I also had to write down, collect, and compile what I did, saw, and discussed. Very concretely, this implied taking a lot of notes. Discussions, meetings, presentations, actions: everything I experienced had, ideally, to be written down, referenced in notebooks and computer documents to be later retrieved, compared, sampled, and analyzed. This full-time data compilation work implied one last move: after my stay within the computer science laboratory—during which I participated in projects, held discussions with colleagues, observed what they did, wrote down as much as I could, and made presentations about my preliminary results (processes that have deeply transformed me and the sociology I now do)—I had to return to my own community of research to more thoroughly work on the collected materials and write an investigation report that, progressively, has become the present book.

但这些过于基础的元素（将在第 1 章中更详细地介绍）却回避了一个重要问题：如何有效地解释、记录和分析计算机科学家在实验室中尝试塑造新算法时所做的事情？如何体验、捕捉和分析他们的行为？

But these all-too-basic elements—that will be more thoroughly presented in chapter 1—elude one important question: How to effectively account for, and thus write down and analyze, what computer scientists do as they try to shape new algorithms within their laboratory? How to experience, capture, and analyze their actions?

行动方针

Courses of Action

一旦人们确信并能够进行实验室研究，以部分但忠实的方式记录算法的构成，人们很快就会进入未知领域。如果有生命科学、物理学、医学或脑科学的实验室研究，很少有研究²⁵进入的成本和开展此类调查所需的时间肯定是造成这种情况的原因之一。但也有可能是一种特殊的思维习惯导致了这种冷漠。事实上，出于一些错综复杂的原因，我将在第 3 章和第 5 章中尝试解决，合理的假设是计算机代码和数学积极地促进了计算机化计算方法的形成，而这个合理的假设往往与不太公平的假设相重叠，即代码和数学都没有或很少有经验厚度。这种假设的算法成分的消逝反过来又使算法显得难以捉摸。这种常见的习惯——Ziewitz (2016) 将其与“算法戏剧” ²⁶联系起来——可能阻止了社会学家进入算法形成、传播和维护的场所：既然一切都发生在那些在那里工作的人的头脑中，为什么还要费心去探究这些地方呢？

As soon as one is convinced of, and enabled to, undertake a laboratory study to document—in a partial yet faithful way—the constitution of algorithms, one quickly lands in uncharted territory. If there are laboratory studies of life sciences, physics, medicine, or brain sciences, very little has been published on computer science work.²⁵ The cost of entry and the time required to carry out this type of investigation certainly contributed to this situation. But it is also possible that a peculiar habit of thought participated in this disinterest. Indeed, for entangled reasons I will try to tackle in chapters 3 and 5, the fair assumption that computer code and mathematics actively contribute to the shaping of computerized methods of calculation is often doubled with the not-so-fair assumption that both code and mathematics have no, or little, empirical thickness. This assumed evanescence of the ingredients of algorithms contributes, in turn, to making them appear inscrutable. This common habit—that Ziewitz (2016) associated with an “algorithmic drama”²⁶—may have discouraged sociologists from entering sites where algorithms are shaped, diffused, and maintained: Why bother trying to inquire into these places since everything happens in the heads of those who work there?

但是，就像任何参与科学实验室日常工作的民族志学者一样——他们努力参与、充分交流并汇编实证材料——我很快意识到，无论同事多么聪明，他们做出的事情都很少。当然，他们从未停止做事——在草稿纸上写字、比较图表、敲击键盘、检查数据库、移动鼠标​​光标、喝咖啡休息——这些事情乍一看似乎毫无关联。但是，当我固执地在日志中记录这些事情时，我很快意识到，这些小的基本“动作块”的连续性有时会形成更大的成就：一个数据库、一个脚本、一个完整的程序、一个算法。通过与我的新同事在他们的实验室里持续呆在一起，认真地记录观察结果，甚至记录一些工作序列（事先得到他们的授权），我很快就被迫承认，我们所说的“实践”实际上是一个没有对立面的术语（Latour 1996）。在我的实验室研究的人工环境中，我尽可能多地解释了各种关联，很快意识到备受争议的“理论”和“实践”之间的区别是一种人为的。在实验室中，只有实践的连续性最终有时会形成“数据库”、“计算机程序”、“数学模型”或“算法”。对这些轨迹的回顾如果准备不足，很容易忽视它们的重要性。但一旦我设法放慢这些轨迹的速度，耐心地解释它们——有时是在那些意识到了它们——我意识到我几乎可以不用任何内部的“抽象”认知机制。

But like any ethnographer involved in the daily work of a scientific laboratory—trying to participate, talk adequately, and compile empirical materials—I quickly realized that very few things could be attributed to the brains of my colleagues, however clever they were. Of course, they never stopped doing things—writing on scratch paper, comparing graphs, typing on keyboards, inspecting databases, moving their mouse cursors, taking coffee breaks—that at first appeared unrelated. But as I stubbornly accounted for these things in my logbooks, I soon realized that the succession of these small elementary “blocks” of action sometimes ended up forming bigger accomplishments: a database, a script, a complete program, an algorithm. By remaining continuously with my new colleagues in their laboratory, conscientiously writing down observations and even recording some work sequences (with their prior authorization), I was soon forced to admit that what we call “practice” is in fact a term without opposite (Latour 1996). In the artificial setting of my laboratory study, accounting for as many associations as possible, I soon realized that the much-debated distinction between “theory” and “practice” was an artifact. In the laboratory, there were only practices whose successions ended up sometimes forming “databases,” “computer programs,” “mathematical models,” or “algorithms.” A little-equipped retrospective look on these trajectories could easily ignore their importance. But once I managed to slow these trajectories down and patiently account for them—sometimes with the help of those who were realizing them—I realized that I could almost do without any internal “abstract” cognitive mechanisms.

继 Jacques Theureau (2003) 的开创性工作之后，我将使用行动过程一词来表示这些可负责任的按时间顺序排列的手势、外表、言语、动作以及人与非人类之间互动的序列，这些序列的表达最终可能会产生某种东西（一块钢铁、一块木板、一项法庭判决、一种算法等）。27^坚持这个通用定义至关重要，因为它有助于我们抵制计算机科学工作的所谓抽象：最终被称为“数学模型”、“代码”甚至“算法”的东西，无论如何，都必须是在特定情况下展开并由可分配的行动者执行的可负责任的行动过程的产物。此外，我将把在不同时间和地点展开但导致相关成就的行动过程纳入通用术语“活动”。在本卷中，活动将被理解为一组相互交织、具有共同最终目的的行动过程。本卷的三个部分都是大胆尝试，旨在展示参与算法形成的活动；因此它们各自的标题都以ing结尾： ground-truth ing、programming ing、formulat ing。

Following the seminal work of Jacques Theureau (2003), I shall use the term courses of action for these accountable chronological sequences of gestures, looks, speeches, movements, and interactions between humans and nonhumans whose articulations may end up producing something (a piece of steel, a plank, a court decision, an algorithm, etc.).²⁷ Sticking to this generic definition is crucial as it will help us resist the supposed abstraction of computer science work: what ends up being called a “mathematical model,” “code,” or even “algorithm” must be, one way or another, the product of accountable courses of action unfolding within specific situations and carried out by assignable actants. Moreover, I shall include under the generic term “activity” courses of action unfolding in different times and locations that yet lead to related achievements. In this volume, an activity will then be understood as a set of intertwining courses of actions sharing common finalities. The three parts of this volume are all adventurous attempts to present activities taking part to the formation of algorithms; hence their respective titles ending with ing: ground-truthing, programming, formulating.

这导致了行动过程的一个潜在限制，因为实验室研究允许对它们进行解释。我之前提到，轨迹必须经常放慢速度，以确定哪些行动过程的表达可能导致某种事物的形成。这种放慢速度是有益的，因为它允许许多关键的塑造行动展开。但它也有一个缺陷：它迫使人们非常缓慢地进行。因此，任何小的先验平凡的行动过程都可能在十几页上展开，从而限制了案例的数量^。28

This leads to one potential limitation of courses of action as laboratory studies allow them to be accounted for. I mentioned earlier that trajectories must often be slowed down to identify the courses of action whose articulation may lead to the formation of something. This slowing down is salutary as it allows many crucial shaping actions to unfold. But it also has one flaw: it forces one to proceed very slowly. As a consequence, any small a priori mundane course of action may unfold on a dozen pages, thus limiting the number of cases.²⁸

三个动名词部分（但可能更多）

Three Gerund Parts (But Potentially More)

我希望读者已经了解我为什么决定进行这项调查，我如何尝试进行调查，以及它最终可能导致的结果。但在深入研究这项探索性研究之前，我将简要介绍这本书的三个部分，按照我的行动导向方法，这三个部分都是动名词：地面实况、编程、制定。

I hope the reader has gotten a sense of why I decided to make this inquiry, how I tried to conduct it, and where it may eventually lead. But before diving in this exploratory study, I shall briefly present the three parts of this book that, following my action-oriented methodology, are all gerunds: ground-truthing, programming, formulating.

第一部分主要讨论定义能够通过计算解决的问题所需的工作。在第一章中，我介绍了总体设置探究并介绍数字图像处理和标准算法研究的基本概念。在第 2 章中，我直接进入问题的核心，并跟踪一群试图发布他们的一种算法的年轻计算机科学家。在第一个图像处理案例研究中，我们将遇到计算机科学家所说的“基本事实”：作为算法材料基础的参考资料库。基本事实的核心地位以及构建它们所需的工作使我断言，在一定程度上，我们得到了基本事实的算法。

Part I mainly deals with the work required to define problems capable of being solved computationally. In chapter 1, I present the overall setting of the inquiry and introduce basic notions in digital image processing and standard algorithmic study. In chapter 2, I go directly to the heart of the matter and follow a group of young computer scientists trying to publish one of their algorithms. During this first case study of image processing in the making, we will encounter what computer scientists call “ground truths”: referential repositories that work as material bases for algorithms. The centrality of ground truths and of the work required to build them make me assert that, to a certain extent, we get the algorithms of our ground truths.

第二部分尝试了一项很少有人尝试过的事情：将计算机编程视为一种实用的情境活动。在第三章中，我提出了编程为什么一直拒绝——并且仍然拒绝——民族志研究的历史和概念原因。在本章的最后，我重点讨论了思维的计算隐喻，这是阻碍对计算机编程实践进行任何仔细分析的主要概念绊脚石。在第四章​​中，我基于前几章中介绍的概念和概念，仔细描述了我在实验室研究期间参加的计算机编程行动课程。除了开辟新的研究途径之外，第二个案例研究还引出了以下命题：程序员可能永远无法解决任何问题。

Part II tries something that has rarely been attempted: considering computer programming as a practical, situated activity. In chapter 3, I propose historical and conceptual reasons why programming has resisted—and still resists—ethnographic scrutiny. At the end of the chapter, I focus on the computational metaphor of the mind, the main conceptual stumbling stone preventing any close analysis of computer programming practices. In chapter 4, building on notions and concepts introduced in the previous chapters, I carefully describe computer programming courses of action I attended during my laboratory study. Besides opening new avenues of research, this second case study leads, inter alia, to the following proposition: a programmer may never solve any problem.

在第三部分中，我考虑了数学在算法形成中的作用。在第 5 章中，我首先基于 STS 启发的数学探究，将数学实践作为科学活动的利益相关者。然后，我使用这种非传统的数学观点将公式化定义为转换实体直到它们获得与先前定义的数学对象相同的形式的活动。在第 6 章中，我基于这些理论论证来解释成功制定地面实况数据库数据之间某些关系的行动方针。这第三个也是最后一个案例研究还将让我们了解地面实况、编程和公式化活动之间的众多联系，这些纠缠在一起的过程有时会导致算法的形成。这些要素最终将使我能够触及机器学习和人工智能的主题，这里认为这是自动化公式化实践的大胆但代价高昂的尝试。在结论中，我提出了这项探究所揭示的经验和理论要素的一些推论。

In part III, I consider the role of mathematics in the formation of algorithms. In chapter 5, I first build on STS-inspired inquiries into mathematics to present mathematical practices as stakeholders of scientific activity. I then use this unconventional view on mathematics to define formulating as the activity of translating entities until they acquire the same form as previously-defined mathematical objects. In chapter 6, I build on these theoretical arguments to account for courses of action that successfully formulated some of the relationships among the data of a ground-truth database. This third and last case study will also make us appreciate some of the numerous links between ground-truthing, programming, and formulating activities, entangled processes that, sometimes, leads to the shaping of algorithms. These elements will finally allow me to touch on the topic of machine learning and artificial intelligence, here considered audacious yet costly attempts at automating formulating practices. In the conclusion, I develop some corollaries of the empirical and theoretical elements this inquiry unfolded.

虽然在本卷中，实地调查、编程和制定活动是相互衔接的，但它们在行动的“真实”生活。在我们很快就会知道的计算机科学实验室等地方，这些活动形成了一个旋风过程，其要素在能动性的舞蹈中相互影响（Pickering 1995）。此外，尽管本书的叙述线索是连续的——后续章节有时会参考前几章——但人们可以用不同的方式浏览它。例如，对民族志叙述感兴趣的读者可能会从一个案例研究跳到另一个案例研究，然后最终回到更理论化的文章，如第 3 章和第 5 章。喜欢概念性探索的读者可能希望反过来，从智力问题开始，然后再回到实际的实践叙述。当然，没有特定期望的好奇读者也可以按照这本书的线索，从第 1 章开始，到结论结束。

Although ground-truthing, programming, and formulating activities follow each other in the present volume, they do not necessarily do so in the “real” life of action. In places such as the computer science laboratory we will soon get to know, these activities form a whirlwind process whose elements influence each other in a dance of agency (Pickering 1995). Moreover, even though this book’s narrative thread is sequential—with subsequent chapters sometimes referring to previous ones—one may browse through it in different ways. Readers interested in ethnographic accounts may, for example, jump from one case study to another before eventually coming back to more theoretical pieces such as chapters 3 and 5. Readers who favor conceptual ventures may wish to go the other way round, starting with intellectual matters before coming back to down-to-earth accounts of practices. Of course, curious readers without specific expectations may also follow the book’s thread, starting from chapter 1 and ending with the conclusion.

如前所述，重要的是要牢记——就像一句咒语——这三种活动构成了算法的经验和党派版本，它们既不是固定的也不是排他性的。尽管我相信它们构成了算法如何产生的令人耳目一新且忠实的概念，但算法的确切生态显然会从进一步的研究中受益。肯定还有更多的活动有助于算法的形成，希望未来的民族志和案例研究能够揭示这些活动。从这个意义上说，尽管本书确实打算提出一种替代的以行动为导向的算法构成，但我的论点也应被视为需要进一步考虑的初步主张。

As mentioned earlier, it is important to keep in mind—almost like a mantra—that these three activities forming an empirical and partisan version of what algorithms shall be are not fixed nor exclusive. Even though they form, I believe, a refreshing and faithful conception of how algorithms come into existence, the precise ecology of algorithms would clearly benefit from further investigations. There are surely many more activities contributing to the formation of algorithms that future ethnographies and case studies will, hopefully, unfold. In that sense, although this volume does intend to bring about an alternative action-oriented constitution of algorithms, my arguments should also be considered preliminary propositions asking for further considerations.

无论如何，铭文只有在阅读时才能创造世界：在这一点上，我主要关心的是读者 - 对算法的构成关系感兴趣的社会学家；对其工作的替代行动导向解释感到好奇的计算机科学家；或者事实上，任何关心算法的力量和美感的人 - 都足够好奇，愿意和我一起探索计算机科学实验室里正在发生的一些事情。

At any rate, inscriptions make worlds only when read: at this point, my main concern is that readers—sociologists interested in the constitutive relationships of algorithms; computer scientists curious about an alternative action-oriented account of their work; or in fact, anyone concerned about the power, and beauty, of algorithms—are intrigued enough to come with me to explore some of the things that are happening in a computer science laboratory.

笔记

Notes

1. 1.过程思想指的是一系列广泛而多样的哲学著作，它们对联想（有时也称为关系）有着相似的感受（Barad 2007；Butler 2006；Dewey [1927] 2016；James [1912] 2003；Latour 1993b, 2013；Mol 2002；Pickering 1995；Serres 1983；Whitehead [1929] 1978）。对于过程思想家来说，正如 Introna 所说（2016，23），“关系不会（因果或其他方式）连接预先存在的实体（或行为者），相反，关系会在生成的流程中制定实体。”事物的本质是它们与其他实体的关联，关联本身是过程的一部分。然后重点放在“如何”而不是“什么”上：过程思维者不会问某物是什么，而是问某物如何变成。因此，这种本体论是关于连续表现而不是二元状态。本书涵盖了这种成为的本体论。

2. 1.  Process thought refers to a wide and heterogeneous body of philosophical works that share similar sensibilities toward associations, sometimes also called relations (Barad 2007; Butler 2006; Dewey [1927] 2016; James [1912] 2003; Latour 1993b, 2013; Mol 2002; Pickering 1995; Serres 1983; Whitehead [1929] 1978). For process thinkers, as Introna put it (2016, 23), “relations do not connect (causally or otherwise) pre-existing entities (or actors), rather, relations enact entities in the flow of becoming.” What things are is what they become in association to other entities, the association itself being part of the process. The emphasis is then put on the “how” rather than the “what”: instead of asking what is something, process thinkers would rather ask how something becomes. This ontology is then about continuous performances instead of binary states. The present volume embraces this ontology of becoming.

3. 2.书的最后附有词汇表，简要定义了本次调查使用的技术术语（例如，行动者、集体世界、宪法、行动方针）。

4. 2.  At the end of the book, a glossary briefly defines technical terms used for this investigation (e.g., actant, collective world, constitution, course of action).

5. 3.这一非传统的社会概念最初由创新社会学中心的玛德琳·阿克里奇、米歇尔·卡隆和布鲁诺·拉图尔发展和推广（Akrich, Callon, and Latour 2006; Callon 1986）。值得注意的是，尽管这一理论观点在学术研究中有所体现，但仍为少数学者所认同。

6. 3.  This unconventional conception of the social has been initially developed and popularized by Madeleine Akrich, Michel Callon, and Bruno Latour at the Centre de Sociologie de l’Innovation (Akrich, Callon, and Latour 2006; Callon 1986). It is important to note that even though this theoretical standpoint has somewhat made its way through academic research, it remains shared among a minority of scholars.

7. 4.拉图尔（2005, 5–6）指出，拉丁语词根socius表示同伴——伙伴——，这与社会性概念相吻合，社会性源于异质实体之间的关联。

8. 4.  As pointed out by Latour (2005, 5–6), the Latin root socius that denotes a companion—an associate—fits well with the conception of the social as what emanates from the association among heterogeneous entities.

9. 5.科学技术研究的异质研究团体的从业者们团结在一起，他们坚信科学不仅仅是逻辑经验主义的表达；世界知识并非预先存在；科学和技术真理依赖于集体安排、仪器和动力（Dear and Jasanoff 2010；Jasanoff 2012）。有关 STS 的全面介绍，请参阅 Felt 等人（2016 年）。

10. 5.  What connects the practitioners of the heterogeneous research community of Science and Technology Studies is the conviction that science is not just the expression of a logical empiricism; that knowledge of the world does not preexist; and that scientific and technological truths are dependent on collective arrangements, instrumentations, and dynamics (Dear and Jasanoff 2010; Jasanoff 2012). For a comprehensive introduction to STS, see Felt et al. (2016).

11. 6.值得注意的是，这种行动能力的下降与依恋社会学无关，依恋社会学正是试图记录令人愉悦的对象的出现，正如 Antoine Hennion (2015, 2017) 所发展的那样。在第 5 章的末尾，我将讨论依恋这一重要概念。

12. 6.  It is important to note that this lowering of capacity to act does not concern the sociology of attachments that precisely tries to document the appearance of delighted objects, as developed by Antoine Hennion (2015, 2017). At the end of chapter 5, I will discuss the important notion of attachment.

13. 7 . 在我看来，“组合”的概念——至少是拉图尔（2010a）所提出的——是广泛使用的“治理”概念的一个优雅的替代品。尽管如此，这两个概念还是有一些共同的特点。首先，这两个概念都假设异质元素组合在一起——人类、机器、物体、公司和机构的集合，试图在同一条船上合作和坚持下去。其次，它们都渴望一个共同的世界，同时接受其各部分的不可简化性：对于这两个概念来说，构成世界的不可简化实体宁愿生活在一个非常知情的社区中，了解不同的、相互竞争的利益，也不愿生活在一个充满不信任和异想天开的荒原上。因此，组合和治理都具有相同的基本研究主题：如何逐步将异质的集体转变为异质的共同世界？第三，它们都同意，传统的集中决策权不再能够实现共同世界的构成；与命令和禁令的垂直性相比，组合和治理更喜欢妥协和谈判的水平性。然而，它们在一个关键点上存在分歧：如果治理仍然承载着对一个平稳但又异质的宇宙的希望，那么组合则促进了对艰苦且不断调整的混乱的需求（Latour 2010a，487）。换句话说，如果控制仍然是治理的一种选择，那么组合则致力于总是令人惊讶的“被迫去做”（Latour 1999b）。正是这种对不断进行创造性调整的需求的强调，让我更喜欢“组合”的概念而不是“治理”。

14. 7.  The notion of “composition”—at least as proposed by Latour (2010a)—is, in my view, an elegant alternative to the widely used notion of “governance.” Both nonetheless share some characteristics. First, both notions suppose heterogeneous elements put together—collectives of humans, machines, objects, companies, and institutions trying to collaborate and persevere on the same boat. Second, they share the desire of a common world while accepting the irreducibility of its parts: for both notions, the irreducible entities that constitute the world would rather live in a quite informed community aware of different and competitive interests than in a distrustful and whimsical wasteland. Both composition and governance thus share the same basic topic of inquiry: how to step-by-step transform heterogeneous collectives into heterogeneous common worlds? Third, they both agree that traditional centralized decisional powers can no longer achieve the constitution of common worlds; to the verticality of orders and injunctions, composition and governance prefer the horizontality of compromises and negotiations. Yet they nonetheless differ on one crucial point: if governance still carries the hope of a smooth—yet heterogeneous—cosmos, composition promotes the need of a laborious and constantly readjusted kakosmos (Latour 2010a, 487). In other words, if control is still an option for governance, composition is committed to the always surprising “made to do” (Latour 1999b). It is this emphasis on the constant need for creative readjustments that makes me prefer the notion of “composition” over “governance.”

15. 8.接下来的两段引自Jaton（2019，319–320）。

16. 8.  The next two paragraphs derive from Jaton (2019, 319–320).

17. 9.自 2000 年代以来，“算法”一词在英美批评文献中变得越来越常见。了解“算法”一词如何取代过去（尤其是 20 世纪 90 年代）也同义使用的其他替代术语（如“软件”、“代码”或“软件算法”）将会很有趣。

18. 9.  The single term “algorithm” became increasingly common in the Anglo-American critical literature from the 2000s onward. It would be interesting to learn more about the ways by which the term “algorithm” has come to take over other alternative terms (such as “software,” “code,” or “software-algorithm”) that were also synonymously used in the past, especially in the 1990s.

19. 10.在Jaton和Vinck（已提交）的论文中，我们仔细考虑了近期算法政治化的具体动态。

20. 10.  In Jaton and Vinck (submitted), we closely consider the specific dynamic of the recent politicization of algorithms.

21. 11.Baya -Laffite、Beaude 和 Garrigues（2018）对这一争议进行了彻底分析。

22. 11.  This controversy has been thoroughly analyzed in Baya-Laffite, Beaude, and Garrigues (2018).

23. 12.正如我们将在本书的实证章节中看到的那样，我们不清楚应该谈论计算机科学家还是工程师。但由于计算机科学的学术领域现在已经很成熟，我选择使用通用术语“计算机科学家”来指代那些每天致力于设计令人惊讶的新算法的人。

24. 12.  As we will see in the empirical chapters of this book, it is not clear whether we should talk about computer scientists or engineers. But as the academic field of computer science is now well established, I choose to use the generic term “computer scientist” to refer to those who work every day to design surprising new algorithms.

25. 13.有关该主题的详细讨论，请参阅 Denis (2018, 83–95)。

26. 13.  For thorough discussions on this topic, see Denis (2018, 83–95).

27. 14.这是否意味着“客观知识”是不可能的？正如我们将在第4、5和6章中看到的那样，得出这样的结论是站不住脚的：尽管科学实践严重依赖的铭文具有不可弥补的局限性，但这些实践仍然能够产生经过证实的客观知识。

28. 14.  Does it mean that “objective knowledge” is impossible? As we will see in chapters 4, 5, and 6, drawing such a conclusion is untenable: despite the irremediable limits of the inscriptions on which scientific practices heavily rely, these practices nonetheless manage to produce certified objective knowledge.

29. 15.在 2004 年的论文中，劳和厄里以哈拉维 (1992, 1997) 最初提出的论点为基础。

30. 15.  In their 2004 paper, Law and Urry build upon an argument initially developed by Haraway (1992, 1997).

31. 16.这在一定程度上解释了一些科学家对 STS 关于“科学事实建构”的工作抱有敌意的反应。关于这个话题，请参阅 Latour (2013, 151–178)。

32. 16.  This partly explains some hostile reactions of scientists regarding STS works on the “construction of scientific facts.” On this topic, see Latour (2013, 151–178).

33. 17.有关最近的例子，请参阅 Cardon (2015) 和 Mackenzie (2017)。

34. 17.  For recent examples, see Cardon (2015) and Mackenzie (2017).

35. 18.在第五章中，我将更详细地讨论科学文献对于形成认证知识的关键重要性。

36. 18.  In chapter 5, I will discuss at greater length the crucial importance of scientific literature for the formation of certified knowledge.

37. 19 . “超常”一词，相对于“超常”而言，最初由 P é rec (1989) 提出。该术语后来被法语社会学所采用，尤其是 Lefebvre (2013)。

38. 19.  The term “infra-ordinary,” as opposed to “extra-ordinary,” was originally proposed by Pérec (1989). The term was later taken up in Francophone sociology, notably by Lefebvre (2013).

39. 20。例如，请参阅 Bishop (2007)、Cormen 等人 (2009)、Sedgewick 和 Wayne (2011)、Skiena (2008) 和 Wirth (1976)。我将在第 1 章中讨论其中一些手册。

40. 20.  See, for example, Bishop (2007), Cormen et al. (2009), Sedgewick and Wayne (2011), Skiena (2008), and Wirth (1976). I will discuss some of these manuals in chapter 1.

41. 21。然而，必须对手册和课程的执行方面保持警惕。金融社会学对这一主题进行了深入研究；例如，参见 MacKenzie、Muniesa 和 Siu (2007) 和 Muniesa (2015)。

42. 21.  However, it is crucial to remain alert to the performative aspects of manuals and classes. This topic is well studied in the sociology of finance; see, for example, MacKenzie, Muniesa, and Siu (2007) and Muniesa (2015).

43. 22 . 这也经常涉及社会科学家采访知名计算机科学家（例如，Seibel 2009；Biancuzzi 和 Warden 2009）。由于这些调查主要关注计算机科学领域备受尊敬的人物，他们的项目取得了很大成功，因此他们的结果往往是回顾性的、总结性的叙述，掩盖了不确定性和脆弱性。关于传记访谈的一些局限性，请参阅 Bourdieu (1986)。关于将民族志简化为访谈的习惯问题，请参阅 Ingold (2014)。

44. 22.  This also often concerns social scientists interviewing renowned computer scientists (e.g., Seibel 2009; Biancuzzi and Warden 2009). As these investigations mainly focus on well-respected figures of computer science whose projects have largely succeeded, their results tend to be retrospective, summarized narratives occluding uncertainties and fragilities. On some limitations of biographic interviews, see Bourdieu (1986). On the problematic habit of reducing ethnography to interviews, see Ingold (2014).

45. 23.有关学者们开始在科学实验室内进行探究的一些原因的介绍，请参阅 Doing (2008)、Lynch (2014) 和 Pestre (2004)。

46. 23.  For a presentation of some of the reasons why scholars started to inquire within scientific laboratory, see Doing (2008), Lynch (2014), and Pestre (2004).

47. 24.关于计算机科学与人文学科（文学、历史、语言学等）之间的和解以及由此产生的数字人文学科的一些有争议但又令人着迷的动态，请参阅 Gold (2012)、Jaton and Vinck (2016) 以及 Vinck (2016)。

48. 24.  On some of the problematic, yet fascinating, dynamics of this rapprochement between computer science and the humanities (literature, history, linguistics, etc.) that gave rise to digital humanities, see Gold (2012), Jaton and Vinck (2016), and Vinck (2016).

49. 25.记录计算机科学工作的少数尝试包括 Bechmann 和 Bowker (2019)、Button 和 Sharrock (1995)、Grosman 和 Reigeluth (2019)、Henriksen 和 Bechmann (2020) 以及 Mackenzie 和 Monk (2004)。我将在本书的实证章节中回顾其中一些研究。

50. 25.  Among the rare attempts to document computer science work are Bechmann and Bowker (2019), Button and Sharrock (1995), Grosman and Reigeluth (2019), Henriksen and Bechmann (2020), and Mackenzie and Monk (2004). I will come back to some of these studies in the empirical chapters of the book.

51. 26 . 在对当代算法批判研究进行全面回顾后，Ziewitz (2016) 警告称，这些研究可能即将陷入一个有问题的僵局。粗略地说，该论点如下：批判研究主要从远处和从其效果的角度考虑算法，因此冒着陷入戏剧循环的风险，不断重复算法之所以强大是因为它们难以捉摸，因为它们强大，因为它们难以捉摸，等等。本书可以看作是某种程度上阻止这种戏剧上演的尝试。在结论中，当我澄清这一探究的政治方面时，我回到了算法戏剧的概念。

52. 26.  After a thorough review of the contemporary critical studies of algorithms, Ziewitz (2016) warned that they could be about to reach a problematic impasse. Roughly put, the argument goes as follows: by mainly considering algorithms from a distance and in terms of their effects, critical studies are taking the risk of being stuck in a dramatic loop, constantly rehashing that algorithms are powerful because they are inscrutable, because they are powerful, because they inscrutable, and so on. The present volume can be considered an attempt at somewhat preventing such a drama from taking hold. In the conclusion, when I clarify the political aspect of this inquiry, I come back to this notion of algorithmic drama.

53. 27. Theureau 的工作在许多方面都是独一无二的。他以法国人体工程学传统（Ombredane 和 Faverge 1955）和对 Newell 和 Simon（1972）认知行为主义以及 Varela 的“行为认知”概念（在第 3 章中讨论）的批判性解读为基础，逐渐提出了一个简单而有效的行动过程定义，即“主体在确定的状态下，积极参与物理和社会确定的环境，并属于确定的文化的可观察活动”（Theureau 2003，59）。他对交通管理（Theureau 和 Filippi 2000）、核反应堆控制（Theureau 等 2001）和音乐创作（Donin 和 Theureau 2007）中涉及的行动过程的分析使他提出了人体工程学研究的“以行动过程为中心的设计”概念。

54. 27.  Theureau’s work is unique in many ways. Building on the French ergonomics tradition (Ombredane and Faverge 1955) and critical readings of Newell and Simon’s (1972) cognitive behaviorism as well as Varela’s notion of “enactive cognition” (discussed in chapter 3), he has gradually proposed a simple yet effective definition of a course of action as an “observable activity of an agent in a defined state, actively engaged in a physically and socially defined environment and belonging to a defined culture” (Theureau 2003, 59). His analyses of courses of action involved in traffic management (Theureau and Filippi 2000), nuclear reactor control (Theureau et al. 2001), and musical composition (Donin and Theureau 2007) has led him to propose the notion of “courses-of-action centered design” for ergonomic studies.

55. 28.在第4章的开始部分，我将简单探讨一下“代表性”的问题。

56. 28.  At the beginning of chapter 4, I will briefly consider the problem of “representativeness.”

实验室

The Lab

实验室位于 CSF 主楼三楼。这是 CSF 的典型组织结构，以正教授（实验室主任）为中心。主任由一名秘书协助处理行政事务，由于合作者来自国外（尤其是波斯、印度和中国）的比例很高，因此行政事务往往很复杂。2 在^这些合作者中，一名博士后学生在实验室待了一年半。一位受邀学者也有一张办公桌，积极参与教学和研究活动。衍生项目成员有时与前面提到的创新领域有关，他们也会在筹集资金期间留在实验室，时间从一到两年不等。这些衍生项目的合作者在实验室研讨会上发表演讲并不罕见（稍后会详细介绍），但在这种情况下，其他合作者必须尊重非官方的“保密协议”。一些处于两份研究合同之间的合作者有时也被聘为“科学家”，这是一个临时职位，允许他们在体面的条件下继续工作。但实验室的大部分成员都是 23 至 30 岁的博士生，通常持有四年的雇佣合同，合同期满后他们需要提交博士论文，从而成为计算机科学博士。我在实验室工作期间，博士生的数量从 6 到 10 不等，取决于提交的论文数量和授予的研究合同。在进行研究活动的同时，这些学生还必须担任本科和硕士课程的助教，包括实验室主任的课程。总而言之，在我合作的两年半里，实验室接待了包括我在内 10 到 16 个人。

The Lab was located on the third floor of the CSF main building. Typical of the organization of the CSF, it was centered upon the tutelary figure of a full professor, the director of the Lab. The director was assisted by a secretary dealing with administrative issues that were often complex due to the high proportion of collaborators who came from abroad (especially from Persia, India, and China).² Among these collaborators, one postdoc student stayed at the Lab for one-and-a-half years. An invited scholar also had a desk and took active part in teaching and research activities. Members of spin-offs, sometimes related to the innovation area mentioned earlier, also stayed within the Lab for the duration of their fund raising, ranging from one to two years. It was not uncommon for these spin-off collaborators to make presentations at Lab seminars (more on this later), though in these situations the other collaborators were required to respect an unofficial “nondisclosure arrangement.” Some collaborators in between two research contracts were also sometimes hired as “scientists,” a temporary position allowing them to pursue their ongoing work in decent conditions. However, most of the Lab’s members were PhD students aged from twenty-three to thirty years old and generally holders of four-year employment contracts, at the end of which they were asked to submit doctoral theses allowing them to become doctors of computer science. During my time in the Lab, the number of PhD students varied from six to ten and depended on the number of submitted theses and awarded research contracts. In parallel to their research activities, these students also had to work as teaching assistants for bachelor’s and master’s classes, including those given by the Lab’s director. All in all, for the two-and-a-half years of my collaboration, the Lab hosted between ten and sixteen people, including myself.

和许多 CSF 教授一样，该主任一直试图在实验室内建立社区动态。例如，在每周实验室会议结束时，她会带上蛋糕和饼干来鼓励大家进行非正式的聊天，在此期间，一两位合作者会介绍他们正在进行的工作。每年还会在附近的餐馆组织两次实验室晚餐；一次是在圣诞节前后，另一次是在六月底。为了响应公司郊游，实验室在夏季还组织了为期两天的旅行。实验室的博士生也为这种动态做出了贡献，他们经常主动组织“下班后”到学校酒吧郊游。所有这些促进工作都有效地建立和维持了合作者之间的关系，其中许多人最初来到实验室时并不认识当地的任何人。

Like many CSF professors, the director continuously tried to establish community dynamics within her Lab. This involved, for example, bringing cakes and biscuits to encourage informal chatting at the end of the weekly Lab meetings, during which one or two collaborators presented their work in progress. Two Lab dinners at nearby restaurants were also organized each year; one around Christmas, the other at the end of June. Echoing a corporate outing, a two-day excursion was organized during the summer as well. The Lab’s PhD students also contributed to this dynamic by frequently organizing “after-work” outings to the school pub on their own initiative. All these facilitation efforts effectively created and maintained relationships among collaborators, many of whom had initially arrived in the Lab without knowing anyone in the area.

在某种程度上，实验室的建筑布局也参与了这些社区动态，因为七个办公室通常由两名研究人员面对面使用，每个办公室都沿着同一个大厅排列（见图1.2和1.3）。实验室还有一个私人自助餐厅，提供桌椅、冰箱和咖啡机。正如我们稍后将看到的，这个自助餐厅经常被用作会议地点，尽管实验室有自己的研讨室。

To some extent, the architectural organization of the Lab also participated in these community dynamics as the seven offices, generally occupied by two researchers facing each other, were each aligned along the same hall (see figures 1.2 and 1.3). The Lab also had a private cafeteria that provided tables, chairs, fridges, and coffee machines. As we will see later, this cafeteria was often used as a meeting point, even though the Lab had its own seminar room.

如果这些社区动态确实有助于创造丰富的工作环境，实验室主任对此大力鼓励，那么它们也与管理方面有关。例如，必须参加实验室会议并做出贡献，每个合作者必须每学期至少做一次演讲。此外，与公司环境类似，合作者生病或丧失工作能力时必须通知秘书，因此建议他们应该在实验室除非另有规定，否则每个工作日都要进行。此外，科研合作者被要求至少每两周与主任会面一次，向她通报他们的研究进展。这让主任能够对正在进行的项目有一个实际的了解，同时让合作者承诺与她分享结果、问题、难题或疑虑。

If these community dynamics, greatly encouraged by the Lab’s director, did contribute to creating an enriching work environment, then they also went along with managerial aspects. For example, attendance and contribution to Lab meetings were mandatory, with each collaborator being required to make at least one presentation per semester. In addition, similar to corporate settings, collaborators were required to inform the secretary in the event of illness or incapacity, thus suggesting they should be at the Lab every working day unless otherwise specified. Moreover, scientific collaborators were asked to meet with the director at least once every two weeks to inform her of their research progress. This allowed the director to have an actualized view on the ongoing projects while committing collaborators to sharing results, questions, problems, or doubts with her.

这让我们看到了贯穿实验室诸多方面的一个核心要素：研究人员被要求产出成果。这种产生切实成果的激励源自更广泛的动态，如今，希望达到并保持世界大学学术排名高位的研究机构普遍存在这种动态（Espeland 和 Sauder 2016）。尽管大多数 CSF 实验室主任都担任稳定的学术职位，但他们仍必须对其研究团队的表现负责，因为对这些评估影响最大的产出类别是发表在同行评议期刊和会议上的文章。我参加和参与的大多数研究工作都朝着这个非常具体的目标前进：发表同行评议文章。尽管实验室与科技行业关系密切，并为衍生产品的推出提供了有效支持，但从这个意义上讲，它主要是以学术论文为导向的。

This leads us to one central element penetrating many aspects of the Lab: researchers were asked to produce outputs. This incentive to produce tangible results derived from a broader dynamic, now common to research institutions desiring to achieve, and maintain, the heights of the academic rankings of world universities (Espeland and Sauder 2016). Although most of the CSF laboratory directors held stable academic positions, they nonetheless had to be accountable for the performance of their research teams as the category of output having the greatest impact on these evaluations were articles published in peer-reviewed journals and conferences. Most of the research efforts I attended and participated in were then directed toward this very specific goal: publishing peer-reviewed articles. Despite its close relations with the tech industry and its effective support for the launch of spin-offs, the Lab was, in that sense, mainly academic-paper oriented.

但是，实验室成员试图在学术期刊和会议论文集上发表的同行评议文章的内容是什么呢？实验室在做什么？实验室的研究领域与一种名为电荷耦合器件 (CCD) 的设备的出现息息相关。CCD 的发展历史，从 20 世纪 60 年代末在贝尔实验室获得专利的概念，到 20 世纪 90 年代支持其工业化的众多规范和标准，是一个漫长而曲折的故事。3^此外，要准确理解其现已稳定的内部功能，需要固体物理学的基础。4^对于我们感兴趣的内容——从表面上了解实验室学术论文的主题——我们可以只关注 CCD 及其不同变体（如互补金属氧化物半导体 (CMOS)）⁵允许实验室做什么（即这些设备所暗示的潜力）。

But what was the content of the peer-reviewed articles that members of the Lab sought to publish in academic journals and conference proceedings? What was the Lab working on? The research field of the Lab was existentially linked to the advent of a piece of equipment called the charge-coupled device (CCD). The history of the CCD’s development, from its patented concept at Bell Labs in the late 1960s to the many norms and standards that supported its industrialization during the 1990s, is a long and tortuous story.³ In addition, a precise understanding of its now-stabilized internal functioning would require foundations in solid-state physics.⁴ For what interests us here—superficially understanding the main topic of the Lab’s academic papers—we can just focus on what CCDs and their different variations such as complementary metal-oxide semiconductors (CMOSs)⁵ allowed the Lab to do (i.e., the potentialities these devices suggest).

简而言之，通过将电磁光子转换为电子电荷以及对其进行放大和数字化，CCD 和 CMOS（作为受许多标准支持的工业生产设备）能够生成由离散方形元素（称为像素）组成的数字图像。6^根据坐标系排列，这些离散像素（分配有像素）允许识别它们在网格中的位置，彩色图像的八位红、绿、蓝值（见图1.4）能够由计算机程序处理，而这些程序本身大多数时候都受到经过认证的数学陈述的启发。前一句中的许多术语将在后面的章节中详细讨论。现在，我们只需了解，在实验室的七个办公室以及许多其他科学和工业场所，建筑物、阴影、山脉、微笑或大象的图片（由标准化 CCD 和 CMOS 拍摄）也被视为二维信号，可以通过计算机计算方法进行处理。7^这些方法的设计和形成、它们在学术论文中的呈现以及它们作为能够自动计算数码照片（通常称为“自然图像”）构成要素的计算机程序的表达是实验室的主要研究重点。⁸这一特定的实践领域过去和现在通常被称为“二维数字信号处理”，或者更简洁地说是“图像处理”或“图像识别”（当它涉及识别任务时）。

In a nutshell, through the translation of electromagnetic photons into electron charges as well as their amplification and digitalization, CCDs and CMOSs—as industrially produced devices supported by many standards—enable the production of digital images constituted of discrete square elements called pixels.⁶ Organized according to a coordinate system allowing the identification of their locations within a grid, these discrete pixels—assigned eight-bit red, green, and blue values in the case of color images (see figure 1.4)—have the ability to be processed by computer programs that are themselves, most of time, inspired by certified mathematical statements. Many terms of the former sentence will be discussed at length in the following chapters. For now, it is enough to comprehend that in each of the seven offices of the Lab as well as in many other scientific and industrial locations, pictures of buildings, shadows, mountains, smiles, or elephants—as produced by standardized CCDs and CMOSs—were also considered two-dimensional signals that could be processed by means of computerized methods of calculation.⁷ The design and shaping of these methods, their presentation within academic papers, and their expression as computer programs able to automatically compute the constitutive elements of digital photographs (often called “natural images”) was the main research focus of the Lab.⁸ This specific area of practice was and is generally called “two-dimensional digital signal processing” or, more succinctly, “image processing” or “image recognition” (when it deals with recognition tasks).

尽管花费时间和精力来构建能够处理 CDD 和 CMOS 像素的计算机计算方法，有意义的方式乍一听可能有些深奥，但这种活动在当代经济中发挥着重要作用。9这与数码照片前所未有的生产、流通和可访问性有关：¹⁰多亏了图像处理算法，这些众多的二维信号已经成为可能表明习惯、属性、偏好和欲望的痕迹。随着图像处理和识别技术^的出现，每天制作和共享的大量数码照片不再是嘈杂、浩瀚的难以捉摸的数据流，而是变成了宝贵的资产（Birch and Muniesa 2020）。必须把握这一现象的重要性。Facebook、谷歌、亚马逊、苹果、IBM 或微软等大型科技服务公司都拥有实验室，其成员每天都在努力制造新算法，以商业化开发数码照片的无限潜力，数码照片是用户、客户和合作伙伴所依附的具体表达。11^民族国家也不例外；强大的公共机构也大量投资于图像处理，以利用数码照片的功能来实现安全、控制和纪律目的。12^近年来，与 Hine (2008) 对生物系统学案例的描述类似，图像处理已被视为控制和规划的资源，并为此日益成为战略政策关注和支持的对象。

Even though spending time and energy assembling computerized methods of calculation capable of processing CDD- and CMOS-derived pixels in meaningful ways might at first sound esoteric, such an activity plays an important role in contemporary economies.⁹ This is to be related with the unprecedented production, circulation, and accessibility of digital photographs:¹⁰ thanks to image-processing algorithms, these numerous two-dimensional signals have become traces potentially indicating habits, attributes, preferences, and desires. Instead of a noisy, expansive stream of inscrutable data, the many digital photographs produced and shared every day have turned into valuable assets (Birch and Muniesa 2020) with the advent of image processing and recognition. This is a phenomenon whose magnitude must be grasped. Giant technology services companies such as Facebook, Google, Amazon, Apple, IBM, or Microsoft all have laboratories whose members work every day to manufacture new algorithms to commercially exploit the infinite potential of digital photographs, tangible expressions of what users, clients, and partners are assumedly attached to.¹¹ Nation-states are not to be left out either; powerful public agencies also massively invest in image processing to make use of the capabilities of digital photographs for security, control, and disciplinary purposes.¹² In recent years, similar to what Hine (2008) described for the case of biological systematics, image processing has been seen as a resource in control and planning and, to this end, has increasingly become the object of strategic policy concern and support.

所有这些听起来可能令人沮丧。然而，图像处理是一个不可避免地令人着迷的研究领域，有许多专门的学术期刊¹³和会议。14^研究问题确实很有吸引力：如何让盒子状的计算机器看到并可能利用它们的形式主义生态，使它们检测、识别和揭示我们作为双足哺乳动物无法用有机感官掌握的事物？每天都有大量的学术努力投入到能够操纵 CCD 和 CMOS 像素的算法的开发中，以使计算机成为真正的视觉设备。然而，值得注意的是，图像处理群体之间的明确界限很难轻易划定：学术研究人员由公共机构资助，也由私人公司资助，而私人公司有时也受到公共机构的邀请，然后参与工业产品的开发。无论好坏，这些异质的行为者相互关联，并通过计算设备合作参与图像处理算法的开发和全球传播。而实验室在其自身层面上也参与了这项高度集体的努力。

All this may sound gloomy. However, image processing is inextricably a fascinating research area with many dedicated academic journals¹³ and conferences.¹⁴ The research issue is indeed appealing: how to make box-like computing machines see and possibly use their formalist ecology to make them detect, recognize, and reveal things that we, as bipedal mammals, cannot grasp with our organic senses? Huge academic efforts are invested every day in the development of algorithms capable of manipulating CCD- and CMOS-enabled pixels to make computers become genuine visual equipment. It is important to note, however, that a clear-cut boundary among image-processing groups cannot be easily drawn: academic researchers are funded by public agencies but also by private companies that themselves are sometimes solicited by public agencies that then take part in the development of industrial products. For better or worse, these heterogeneous actants associate with each other and cooperatively participate in the development and worldwide diffusion of image-processing algorithms through computing devices. And at its own level, the Lab was participating in this highly collective endeavor.

然而，有人可能会反对说，像实验室这样的 16 人图像处理学术实验室并不等同于谷歌这样的大型技术服务公司或国家安全局这样的强大政府机构。我怎么敢把一个小而受人尊敬的学术机构与一个对算法制造感兴趣的民族志学者和对保密性有浓厚兴趣、每天都在为逐步建立“黑箱社会”做出贡献的巨型参与者放在同等水平上？（Pasquale 2015）确实，作为学术命题的算法与作为商业产品或实际控制设备的算法之间存在重要差异（特别是在优化和软件实现方面）。然而，必须指出的是，像实验室这样的学术贡献确实为大型工业和国家参与者的工作提供了支持。这些联系通常在内部会谈中显现出来，在会谈中，工业界的校友被邀请在学术环境中讨论他们正在进行的项目。在实验室期间，我参加了许多这样的讲座。起初，我惊讶地发现，在 Google Brain 或 IBM Watson 等令人印象深刻的先验关系背后，隐藏着一位与我日常接触的计算机科学家并无太大不同的计算机科学家，他们说着大致相同的事情，在规模相似的团队中工作（尽管薪水截然不同）。例如，2015 年 11 月，实验室主任邀请了一位 Instagram 员工（实验室的校友）介绍他们的新浏览系统，该系统的主要组件源自2014 年 IEEE 计算机视觉与模式识别会议论文集上发表的一篇论文。2014 年 6 月，一位在 NEC 工作的前实验室成员在一个五人团队中介绍了她正在进行的算法项目，该项目源自她参加的 2013 年欧洲计算机视觉会议上发表的一系列论文。其他人（主要来自 IBM 和谷歌）也参加了实验室和邻近的 CSF 信号处理实验室组织的这些“受邀演讲”，大多数时候都会提到并使用最先进的出版物。15^正式成为行业参与者的参与者似乎与学术界联系密切，他们在类似规模的团队中工作，参加相同的活动，并分享相同的参考文献。更妙的是，实验室等学术实验室与庞大的科技行业之间的这种持续互动是双向的：谷歌、Facebook 和微软等公司也组织学术活动、赞助国际会议并在排名靠前的期刊上发表论文（见图1.5）^。16

Yet one may rightly object that a sixteen-person academic laboratory for image processing such as the Lab is not akin to, say, a giant technology services company such as Google or a powerful state agency such as the National Security Agency. How dare I treat on the same level a small yet respected academic institution welcoming an ethnographer interested in the manufacture of algorithms and gigantic actors attached to secrecy and daily contributing to the progressive establishment of a “black box society” (Pasquale 2015)? It is true that important differences exist between an algorithm as an academic proposition and an algorithm as a commercial product or an actual control device (notably in terms of optimization and software implementation). Nevertheless, it is crucial to specify that academic contributions such as those of the Lab do irrigate the work of large industrial and state actors. These connections are often made visible during in-house talks where alumni working in the industry are invited to discuss their ongoing projects in academic settings. During my stay at the Lab, I attended many such talks and was at first surprised to find that behind a priori impressive affiliations such as Google Brain or IBM Watson lay a computer scientist not so dissimilar to the ones I daily interacted with, saying more or less the same things, and working in teams of similar proportions (though for a significantly different salary). For example, in November 2015, the director of the Lab invited an Instagram employee—an alumnus of the Lab—to talk about their new browsing system whose main components derived from a paper published in the Proceedings of the 2014 IEEE Conference on Computer Vision and Pattern Recognition. In June 2014, a former Lab member working for NEC in a five-person team also presented her ongoing algorithmic project as deriving from a series of papers presented at the 2013 European Conference on Computer Vision in which she participated. Other people—mostly from IBM and Google—also took part in these “invited talks” organized by the Lab and neighboring CSF signal-processing laboratories, most of the time mentioning and using state-of-the-art publications.¹⁵ Actors who were officially part of the industry appeared then closely connected to the academic community, working in teams of similar size, participating in the same events, and sharing the same references. Better still, this continuous interaction between academic laboratories such as the Lab and the gigantic tech industry was a two-way street: companies like Google, Facebook, and Microsoft also organized academic events, sponsored international conferences, and published papers in the best-ranked journals (see figure 1.5).¹⁶

尽管如此，学术出版物不是商业产品，这一点依然是事实；如果大学和工业实验室都发表论文介绍新的图像处理算法，那么这些方法很少能按原样工作。要成为能够在集体世界中产生重大影响的真正商品，它们必须参与更广泛的钝化和估值过程，这将显著改变它们的初始属性（Callon 2017；Muniesa 2011b）。根据它们在差异化网络中的流通情况，一些最初由工业或学术图像处理实验室设计的计算机化计算方法因此可以保持非常专业化并用于临时目的（例如超像素分割算法），而其他方法则可以在更广泛的组合中广泛应用并在工业上实施，例如数码相机（例如红眼消除算法）、昂贵的软件和大型信息系统（例如文本识别算法、压缩方案或特征聚类）。然而，在图像处理算法在更广泛的网络中传播并混合成为更大系统的一部分之前，它们首先需要在异构研究社区中进行设计、讨论和共享，实验室在其中发挥了积极作用。无论是广泛使用的还是专门的，图像处理算法（在计算机科学界有时也被称为“模型”）首先需要在实验室等地方进行培育、训练、评估和比较。

Nonetheless it remains true that academic publications are not commercial products; if university and industrial laboratories both publish papers presenting new image-processing algorithms, then these methods are rarely workable as they are. To become genuine goods capable of making important differences in the collective world, they must take part in wider passivation and valuation processes that will significantly modify their initial properties (Callon 2017; Muniesa 2011b). Depending on their circulation within differentiated networks, some computerized methods of calculation initially designed by industrial or academic image-processing laboratories can thus remain very specialized and intended for ad hoc purposes (e.g., superpixel segmentation algorithms), whereas others can become widespread and industrially implemented in broader assemblages such as digital cameras (e.g., red-eye-removal algorithms), expensive software, and large information systems (e.g., text-recognition algorithms, compression schemes, or feature clustering). However, before they may circulate in broader networks and hybridize to the point of becoming parts of larger systems, image-processing algorithms first need to be designed, discussed, and shared among a heterogeneous research community in which the Lab played an active role. Whether widespread or specialized, image-processing algorithms—also sometimes just called “models” within the computer science community—first need to be nurtured, trained, evaluated, and compared in places like the Lab.

因此，开发图像处理算法并将其发表在同行评审的学术期刊和会议上是实验室的核心活动，而我打算记录的正是这项活动。但我仍然必须找到一种方法来记录在那里发生的行动过程。

Developing image-processing algorithms and publishing them in peer-reviewed academic journals and conferences was thus a central activity within the Lab, and it was this activity that I intended to account for. Yet I still had to find a way to document the courses of action that took place there.

收集材料

Collecting Materials

由于我的跨学科研究合同，我在实验室工作了两年半。和其他合作者一样，我有一张办公桌、一个电子邮件地址和一个行政系统内的账户。然而，尽管这些是民族志研究的最佳条件，但说最初的几天很艰难也不过分：我周围发生的一切起初似乎都遥不可及。幸运的是，我必须遵守的实验室规则让我很快就能体验到可分配的情况。我将这些情况逐步分为七种不同但相互关联的情况这些类型的系统描述和引用最终构成了我的现场数据语料库。

Thanks to my interdisciplinary research contract, I was part of the Lab for two-and-a-half years. Just as any other collaborator, I had a desk, an e-mail address, and an account within the administrative system. Yet despite these optimal conditions for ethnographic investigation, it would be an understatement to claim that the first days were difficult: everything happening around me seemed at first out of reach. Fortunately, the rules of the Lab that I had to observe quickly allowed me to experience assignable situations. I divided these situations progressively into seven different yet interrelated types whose systematic account and referencing ended up constituting my corpus of field data.

我遇到的第一类情况是我之前提到的实验室会议。在这些每周的会议中，实验室成员聚集在一个小的会议室里，参加并回应正在进行的工作的演示。每个博士生（包括我）、博士后、衍生成员或受邀学者都被要求每学期至少做一次演讲。这些会议对我的研究至关重要，至少有三个原因。首先，它们帮助我确定了新同事的研究主题。然后，我可以利用这些信息在更非正式的环境中与他们展开讨论。其次，实验室会议让我能够在整个实验室面前展示我的研究项目及其一些初步建议。这些强制性的练习迫使我把我的探索直觉付诸实践，并经常对其进行改进。第三，这些情况让我有机会分享疑虑和需求，就像 2015 年 9 月，我利用这个论坛公开寻求帮助，试图更好地记录计算机编程实践（第 4 章将对此进行更多介绍）。然而，尽管这些实验室会议对于我的研究进展至关重要，但接下来章节中我将使用的大部分数据并非在这些情况下收集的。事实上，由于这些会议主要讨论实验室内正在进行的研究项目的结果，因此导致这些结果的实证过程和行动方针通常不是讨论的中心。

The first type of situation I experienced was the Lab meetings I mentioned earlier. During these weekly meetings, the Lab’s members gathered in a small conference room to attend and react to presentations of works in progress. Every PhD student (me included), postdoc, spin-off member, or invited scholar were asked to make at least one presentation each semester. These meetings turned out to be crucial to my inquiry for at least three reasons. First, they helped me identify the research topics of my new colleagues. I could then use this information to initiate discussions with them in more informal settings. Second, Lab meetings allowed me to present my research project as well as some of its preliminary propositions in front of the whole Lab. These mandatory exercises thus forced me to put my exploratory intuitions to the test and, often, retrofit them. Third, these situations gave me opportunities to share doubts and needs as in September 2015 when I used this tribune to publicly ask for help in my attempts to better document computer programming practices (more on this in chapter 4). Yet although these Lab meetings were essential to the advancement of my inquiry, most of the data I will use in the following chapters were not collected during these situations. Indeed, as these meetings mostly dealt with results of ongoing research projects within the Lab, the empirical processes and courses of action that led to these results were generally not at the center of the discussions.

第二种情况是实验室和邻近信号处理实验室组织的会议。如前所述，其中一些会议是受邀演讲，业内的校友前来讨论正在进行的项目。其他会议更接近传统的主题演讲，主要由来自学术机构的知名研究人员发言。虽然我再次没有在实证章节中直接使用从这些会议收集的数据，但这些事件仍然是需要体验和考虑的重要情况，因为它们让我能够识别计算机科学中的当前争论，并更好地理解研究和行业之间的一些关系。

The second type of situation was conferences organized by the Lab and neighbored signal-processing laboratories. As mentioned earlier, some of these conferences were invited talks where alumni working in the industry came to discuss ongoing projects. Other conferences were closer to traditional keynotes and gave the floor to prominent researchers, mainly from academic institutions. Though, again, I do not directly use data collected from these conferences in the empirical chapters, these events were nonetheless crucial situations to experience and account for as they allowed me to identify current debates in computer science and better appreciate some of the relationships between research and industry.

我经历的第三种情况是所谓的小组会议，我在 2013 年 11 月至 2014 年 6 月期间参加了这些会议。这些小组会议是实验室主任指派给我的一个图像处理项目的一部分，对我的民族志研究非常有价值。探究的过程中，它们让我遇到了计算机科学家所说的基本事实——不显眼但对算法的形成至关重要的实体。这些实体将在第 2 章中介绍，并将贯穿本书的其余部分。

A third type of situation I experienced was the so-called Group meetings in which I participated between November 2013 and June 2014. These Group meetings were part of an image-processing project to which the Lab’s director had assigned me, and they were precious for my ethnographic inquiry as they made me encounter what computer scientists call ground truths—inconspicuous entities that are yet central to the formation of algorithms. These entities will be introduced in chapter 2 and will accompany us throughout the rest of the book.

第四种情况发生在实验室的办公桌上。找到适当的方法来解释这些“办公桌情况”是本研究的一个重要条件，因为正是在这些精确的时刻和地点，对算法的实际构建至关重要的行动过程经常发生。在 2013 年 11 月至 2014 年 6 月期间我被分配到的图像处理项目的一小部分中（第 6 章中有更多内容），以及在 2015 年 9 月至 2016 年 2 月期间发生的几次计算机编程事件中（第 4 章中有更多内容），我有机会跟踪和解释此类办公桌情况。

A fourth type of situation took place at the office desks of the Lab. Finding appropriate ways to account for these “desk situations” was an important felicity condition of this inquiry as it was at these precise moments and locations that courses of action crucial to the actual construction of algorithms often took place. I had the chance to follow and account for such desk situations during a small part of the image-processing project to which I was assigned between November 2013 and June 2014 (more on this in chapter 6) as well as during several computer programming episodes that took place between September 2015 and February 2016 (more on this in chapter 4).

第五种情况是我在实验室期间参加的众多课程和教程。从基础信号处理课程到高级 Python 编程教程，我的大部分时间和精力都用于学习计算机科学语言。即使我没有在以下案例研究中直接使用我在课堂或教程中看到的元素，这些情况仍然极大地帮助了我与计算机科学家同事的交流。虽然相当耗时——同样，我最初没有计算机科学经验——但这些学习活动是与我的同事就他们关心的问题进行充分互动的关键先决条件。

A fifth type of situation was the numerous classes and tutorials in which I participated throughout my time at the Lab. From basic signal-processing classes to advanced Python programming tutorials, a significant part of my time and energy was dedicated to learning the language of computer science. Even if I do not directly use elements I saw in classes or during tutorials in the following case studies, these situations nonetheless greatly helped me speak with my computer scientist colleagues. Though quite time consuming—again, I had initially no experience in computer science—these learning activities were crucial prerequisites to interact adequately with my fellow workers about issues that mattered to them.

第六种情况是我在实验室期间进行的半结构化访谈。这些访谈最初是探索性的，旨在让我更好地了解同事如何看待他们的工作。然而，随着调查的进展，我转而使用访谈作为追溯工具，与实验室成员一起回顾那些我只能部分解释的事件。这帮助我填补了数据中的许多空白。

A sixth type of situation was the semi-structured interviews I conducted throughout my stay at the Lab. These interviews were initially exploratory in nature and aimed to give me a better understanding of how my colleagues saw their work. However, as the investigation progressed, I instead used interviews as retroactive tools to revisit with Lab members the events for which I could only partially account. This helped me fill in some of the many gaps in my data.

最后，第七种常见的情况是我每天与实验室成员进行的非正式讨论。虽然我进行了 25 次半结构化访谈，但这些访谈显然不如我在咖啡休息、午餐、圣诞派对、公司郊游或下班后在酒吧进行的无数次对话那么有价值。除了促进我融入实验室之外，这些情况还帮助我分享我所经历和记录的内容。在这些非正式时刻，我可以讨论过去的演讲、最近发表的论文、正在进行的项目、即将进行的编程操作，或者我在课堂上看到的不清楚的元素。

Finally, a seventh generic type of situation was the informal discussions I had daily with the Lab’s members. Although I conducted twenty-five semi-structured interviews, these were clearly not as valuable as the numerous conversations I had during coffee breaks, lunches, Christmas parties, corporate outings, or after-work sessions at the pub. Besides facilitating my integration within the Lab, these situations helped me share what I was experiencing and documenting. During these informal moments, I could, for example, discuss past presentations, recently published papers, ongoing projects, forthcoming programming operations, or unclear elements I had seen in class.

从 2013 年 11 月到 2016 年 4 月，我大部分工作时间都在实验室内外度过，在这七种情况之间切换，并尽可能将它们记录在我的日志中。一天结束时，有时直到深夜，我都会使用文本编辑器清理这些笔记，根据越来越一致的分类法对它们进行分类，并将它们与它们所来自的纸质页面进行引用（见图1.6）。这种收集和引用系统起初非常混乱，因为情况类别的数量增加到不再相关的程度，而我最初的单个 Word 文档变得越来越笨重。然而，几个月后，我可以识别出我刚刚提出的七种不同但相互关联的情况类别，而且由于我通过课程和教程逐渐掌握了计算机编程技能，我决定坚持使用单独的 .txt 文件，这些文件的内容可以通过我开始起草的简单但功能强大的 Python 程序浏览（见图1.7）。一旦系统化，这个临时的数据管理计划或多或少可以让我灵活地处理我的数字化数据，同时保持对原始纸质笔记的访问。

From November 2013 to April 2016, I spent most of my working time in and around the Lab, switching among these seven types of situations and trying to account for them in my logbooks the best I could. At the end of the day, sometimes until late in the evening, I used a text editor to clean up these notes, classify them according to an increasingly consistent taxonomy, and reference them to the paper pages from which they derived (see figure 1.6). This collecting and referencing system was at first very messy as the number of situational categories increased to the point of no longer being relevant and my single initial Word document became increasingly cumbersome. However, after a couple of months, I could identify the seven different yet interrelated situational categories I have just presented, and thanks to the computer programming skills I progressively acquired through classes and tutorials, I decided to stick to individual .txt files whose content could be browsed by simple yet powerful Python programs I started to draft (see figure 1.7). Once systematized, this ad hoc data management plan more or less nimbly allowed me to juggle my digitized data while maintaining access to the original paper notes.

2016 年 4 月，在举行了一场小型的告别派对后，我离开了实验室，带走了大约一千页手写笔记、两千个 .txt 文件、十几个可模块化的 Python 脚本、数百个音频、图像和电影录音以及大量未完成的分析命题。带着所有这些实证材料，我（暂时）离开了我的实地考察点，认真思考这一切的意义。

In April 2016, after a small farewell party, I left the Lab with around one thousand pages of handwritten notes; two thousand .txt files; a dozen modulable Python scripts; and hundreds of audio, image, and movie recordings as well as numerous half-finished analytical propositions. And with all these empirical materials literally under my arm, I (temporarily) exited my field site, asking myself serious questions about the significance of all this.

痛苦的插曲

A Torturous Interlude

民族志是一种变革性的体验。接触世界并写下它们——尝试这种奇怪的练习有什么意义？计算机科学现在让我感到安慰。至于我以前的社会学家同事，他们会怎么看待这个新的我？我再也无法说话了。这是一段艰难的旅程，意义重大的蜕变：“我明白，既然我无法用异教徒的语言表达自己，我宁愿保持沉默，”很久以前有人说过。然而，文字必须写下来，承诺必须兑现，有一件事不能忘记：我的新“新”同事（以前的同事）都经历了类似的旅程。毕竟，我们处于同样不稳定的船上，试图从零散的经验数据中写出忠实的社会学文献。但是，我怎样才能公正地对待我有限而又经验性的材料，那些我提议成为其代言人（没有任何授权）的人的扭曲声音？我什么都缺：历史、媒介、语言。我从哪里开始？也许像往常一样，从事情的中间开始。回到基础，回到实践，回到行动方针。阅读和重读经典；一次又一次地深入研究我的材料，同时与我的同事分享，他们逐渐又成为了搭档（我怎么会忘记这一点？）。半相关的东西开始出现——几乎是分析性的命题。什么数据能让它们在书面文件中绽放？甚至不是一小部分，是一个无限小的数量：一个开明世界的微小快照。负责任的活动开始在文本页面上成形。但它们仍然可读吗？铭文只有在阅读时才能创造世界。概念短缺：计算机科学和社会学可能都没有办法对抗算法的制造。最细微的编程序列很快就暗示着计算机历史的改写；任何小公式都需要一种另类的数学哲学（这是一个多么混乱的话题！）。我们闭着眼睛四处走动。然而，逐渐地，模式出现了：行动路线成为追踪真实、负责任的活动的载体；一个印象派的草稿，从中出现了对抗性的线条：它们可能很强大，但并非不可捉摸。我们如何开始用算法创作？希望如此渺茫，手段如此有限。“沙漠中传来一声呼喊”，等等。足够的哀叹：整个事情都是由问题驱动的比我个人的小麻烦更重要。我想我现在必须验证我的回程票，以某种方式提出一个部分但经验性的算法构成。

Ethnography is a transformative experience. Encountering worlds and writing about them—what is the point of even trying such an odd exercise? Computer science now gives me comfort. And as for my former sociologist peers, what will they think of this new me? I cannot talk anymore. Hell of a journey, significant metamorphosis: “I understand, and since I cannot express myself except in pagan terms, I would rather keep quiet,” someone said a long time ago. Yet words shall be written, promises kept, and something not forgotten: my new “new” colleagues (the former ones) have all gone through similar journeys. After all, we are in the same shaky boat, trying to write faithful sociological documents from scattered empirical data. But how can I do justice to my limited yet empirical materials, distorted voices of those for whom I proposed to become the spokesperson (without any mandate)? I lack everything: a history, a medium, a language. Where do I start? Maybe in the middle of things, as always. Back to fundamentals, to practices, to courses of action. Read and reread classics; dive again and again into my materials while sharing them with my colleagues who are gradually becoming pairs again (how could I have forgotten that?). Half-relevant things start to emerge—almost-analytical propositions. What data can make them bloom in a written document? Not even a fraction, an infinitesimal quantity: tiny snapshot of an enlightened world. Accountable activities start taking shape on text pages. But are they still readable? Inscriptions only make worlds when read. Conceptual shortage: both computer science and sociology may not have the means to confront the manufacture of algorithms. The slightest little programming sequence soon suggests the rewriting of computers’ history; any small formula demands an alternative philosophy of mathematics (what a cluttered topic!). We walk around with eyes wide shut. Gradually, though, patterns emerge: courses of action become vectors tracing genuine, accountable activities; an impressionist draft from which adversarial lines appear: they may be powerful but not inscrutable. How could we start composing with algorithms? The hope is so dim, and the means so limited. “A voice cries out in the desert,” and so on and so on. Enough laments: the whole thing is driven by issues more important than my small personal troubles. And I guess I must now validate my return ticket to propose a partial-yet-empirical constitution of algorithms, somehow.

您说的是算法吗？

Algorithm, You Say?

通过前面不寻常的部分，我希望读者能够理解，撰写有关算法形成的民族志文献可能有些曲折——当人们意识到在计算机科学中算法的概念很少成问题时，就更加如此！作为一名对算法制造感兴趣的社会学家和民族志学者，我确实进入了一个学术领域，这个领域的最杰出人物已经——并且仍然——毕生致力于算法研究。对于许多计算机科学专业人士来说，“算法是什么”的争论被夸大了；正如一位同事在我进入实验室的第一周建议的那样，参加当地的“算法研究”本科课程可能会让我在创纪录的时间内完成我的研究……为了规范我的分析姿态，因此，重要的是要看看这种成熟的面向计算机科学的算法，将本研究视为对它的原始补充。

Going through the previous, unusual section, I hope the reader could appreciate that writing an ethnographic document about the shaping of algorithms can somewhat be tortuous—even more so when one realizes that in computer science the notion of algorithm is rarely problematic! As a sociologist and ethnographer interested in the manufacture of algorithms, I indeed landed in an academic field whose most illustrious figures have dedicated—and still dedicate—their lives to the study of algorithms. To many computer science professionals then, the fuss about “what an algorithm is” is overhyped; as one colleague suggested me on my first week in the Lab, taking the local undergraduate course in “algorithmic study” may allow me to complete my research in record time… In order to specify my analytical gesture, it is thus important to look at this well-established computer-science-oriented take on algorithms to consider the present work as an original complement to it.

浏览大量（但不是无限的）关于算法研究的计算机科学手册时，我们会注意到算法的定义相当同质化。作者通常首先简要介绍一下这个术语的历史，¹⁷然后迅速转到其当代普遍的定义，即由不同步骤组成的系统方法。18^然后，作者指定算法步骤的规则应该足够明确，以便在计算设备中实现，从而将算法与其他先验系统方法（如烹饪食谱或安装指南）区分开来。同时，还指定这些逐步的计算机可实现方法始终指向它们要解决的问题。19^第二个定义元素为算法分配了一个功能，使计算机能够提供相对于手头的特定问题正确的答案。

When browsing through the numerous—yet not infinite—computer science manuals on algorithmic study, one notices algorithms are defined in quite a homogeneous way. Authors typically start with a short history of the term¹⁷ before quickly shifting to its general contemporary acceptation as a systematic method composed of different steps.¹⁸ Authors then specify that the rules of an algorithm’s steps should be univocal enough to be implemented in computing devices, thus differentiating algorithms from other a priori systematic methods such as cooking recipes or installation guides. In the same movement, it is also specified that these step-by-step computer-implementable methods always refer to a problem they are designed to solve.¹⁹ This second definitional element assigns algorithms a function, allowing computers to provide answers that are correct relative to specific problems at hand.

在这些开场白之后，计算机科学手册倾向于围绕“输入”和“输出”来组织这些功能性、一步一步的计算机可实现的问题解决方法。因此，算法的功能活动被进一步指定：算法可以提供定义问题的正确答案是将输入转换为输出。这第三次定义运动导致了算法的标准概念，即“一种将任何可能的输入实例转换为所需输出的过程”（Skiena 2008，3）^。20

Right after these opening statements, computer science manuals tend to organize these functional step-by-step computer-implementable problem-solving methods around “inputs” and “outputs.” The functional activity of algorithms is thus further specified: the way algorithms may provide right answers to defined problems is by transforming inputs into outputs. This third definitional movement leads to the standard well-accepted conception of algorithm as “a procedure that takes any of the possible input instances and transforms it to the desired output” (Skiena 2008, 3).²⁰

这些先验的、过于基本的元素实际上并不简单，因为它们以评估立场推动着算法的发展，并以一种非常有针对性的方式构建算法。事实上，通过赋予算法问题输入和解决方案输出，这种对算法的看法可以强调这两极之间的充分性关系。算法的研究变成了对其有效性的研究。这种忽视的立场是根本性的，渗透到了整个算法研究领域，Knuth 很好地总结了该领域的科学议程：“我们经常面临针对同一个问题的几种算法，我们必须决定哪种是最好的”（1997a，7；斜体添加）。21^从这一点来看，算法分析可以专注于制定元方法，从而系统化算法的形式评估。

These a priori all-too-basic elements are, in fact, not trivial as they push ahead with an evaluation stance and frame algorithms in a very oriented way. Indeed, by endowing itself with problems-inputs and solutions-outputs, this take on algorithms can emphasize on the adequacy relation between these two poles. The study of algorithms becomes then the study of their effectiveness. This overlooking position is fundamental and penetrates the entire field of algorithmic study whose scientific agenda is well summarized by Knuth: “We often are faced with several algorithms for the same problem and we must decide which is best” (1997a, 7; italics added).²¹ From this point, algorithmic analyses can focus on the elaboration of meta-methods that allow the systematization of the formal evaluation of algorithms.

算法研究者提出的算法分析方法借鉴了各种数学分支（例如集合论、复杂性理论），非常优雅和强大。此外，鉴于实施、数据结构、优化和理论理解方面的重大进步，将算法视为输入和输出之间或多或少功能性的接口（它们本身由特定问题定义）的标准概念当然值得高度重视。然而，我认为这种标准概念有一些局限性，在算法争议不断的今天，这些局限性足以提出仍需提交的补充替代方案。

Borrowing from a wide variety of mathematical branches (e.g., set theory, complexity theory), methods for analyzing algorithms as proposed by algorithmic students can be extremely elegant and powerful. Moreover, in the light of the significant advances in terms of implementation, data structuration, optimization, and theoretical understanding, this standard conception of algorithms as more or less functional interfaces between inputs and outputs—themselves defined by specific problems—certainly deserves its high respectability. However, I believe this standard conception has some limits that, in these days of controversies over algorithms, are important enough to suggest complementary alternatives that yet still need to be submitted.

首先，算法的标准概念忽视了算法旨在解决的问题的定义。根据这种观点，问题及其潜在解决方案已经存在，算法研究的作用是评估将输入转化为输出的步骤的有效性。然而，可以公平地假设问题和定义它们的术语本身并不存在。例如，正如本书第 2 章所示，问题是问题化过程精心灌溉的产物，涉及习惯、愿望、技能和价值观。这些集体过程极大地参与了算法（作为解决问题的设备）的进一步设计方式。

First, the standard conception of algorithms overlooks the definition of the problems that algorithms are intended to solve. According to this view, problems and their potential solutions are already made, and the role of algorithmic studies is to evaluate the effectiveness of the steps leading to the transformation of inputs into outputs. Yet it is fair to assume that problems and the terms that define them do not exist by themselves. As it is shown in chapter 2 of this book, for example, problems are delicately irrigated products of problematization processes engaging habits, desires, skills, and values. And these collective processes greatly participate in the way algorithms—as problem-solving devices—will further be designed.

第二个限制与第一个限制相关：如果我们将问题化视为算法设计的一部分，那么算法之间的竞争性质算法在不断变化。最好的算法不仅是那些形式特征证明其优越性的算法，而且是那些设法将问题定义与能够评估其结果的程序联系起来的算法。通过专注于形式标准（而不考虑这些形式主义如何参与手头问题的初步形成），算法的标准概念往往会掩盖算法的评估基础设施和政治。例如，如第 2 章所示，评估程序不一定遵循算法的设计；它们有时也先于并影响算法的设计。

The second limit is linked to the first one: if one considers problematization as part of algorithmic design, the nature of the competition among algorithms changes. The best algorithms are not only the ones whose formal characteristics certify their superiority but also the ones that managed to associate with their problems’ definitions the procedures capable of evaluating their results. By concentrating on formal criterions—without taking into account how these formalisms participated in the initial shaping of the problems at hand—the standard conception of algorithms tends to cover up the evaluation infrastructure and politics of algorithms. As shown in chapter 2, for example, evaluative procedures do not necessarily follow the design of algorithms; they also, sometimes, precede and influence it.

第三，没有考虑迭代方法的实际计算机化。尽管算法的标准概念正确地坚持计算机代码对于算法的最佳执行具有核心作用，但这种坚持以编程方法的形式出现，而这些方法并没有考虑在计算机终端上进行的编程体验。根据算法的这种标准概念，编写能够以所需方式触发电脉冲的编号指令列表主要被视为一种达到目的的手段。但正如本书第 4 章和第 6 章所示，编程实践——凭借其展开所需的集体过程——有时也会影响算法的产生方式。

Third, the actual computerization of the iterative methods is not considered. Even though the standard conception of algorithms rightly insists on the centrality of computer code for the optimal execution of algorithms, this insistence takes the shape of programming methodologies that do not consider the experience of programming as it is lived at computer terminals. According to this standard conception of algorithms, writing numbered lists of instructions capable of triggering electric pulses in desired ways is mainly considered a means to an end. But as it is shown in chapters 4 and 6 of this book, programming practices—by virtue of the collective processes they require in order to unfold—also sometimes influence the way algorithms come into existence.

第四，关于数学语句最终如何被纳入到输入到输出的转换中，以及这种纳入如何影响所考虑的算法，我们谈得很少。对于算法的标准概念，数学语句是突然出现的，随时可以通过其他能够评估其有效性的数学语句进行审查。然而，正如本书第 6 章所指出的，为了将输入转化为输出而纳入数学语句本身就是一个有问题的过程，而且这又会影响算法的性质。数据集的初始概念及其逐步的问题化、重组和缩减涉及充分参与野外算法生态的期望和预期。

Fourth, little is said about how mathematical statements end up being enrolled for the transformation of inputs into outputs and how this enrollment affects the considered algorithms. To the standard conception of algorithms, mathematical statements appear out of the blue, ready to be scrutinized by means of other mathematical statements capable of evaluating their effectiveness. Yet as the chapter 6 of this book indicates, enrolling mathematical statements in order to operate the transformation of inputs into outputs is a problematic process in its own right, and again, this impacts the nature of algorithms. The initial conception of the dataset and its progressive problematization, reorganization, and reduction engage expectations and anticipations that fully participate in the ecology of algorithms in the wild.

因此，本研究旨在开拓算法，并将其扩展到与算法相关但标准概念无法理解的流程。当然，如果这一尝试的目的不是质疑算法研究的结果，那么它的目的是用扎实的社会学考虑来丰富算法研究。

The present work therefore intends to open up algorithms and extend them to processes that they are attached to but whose standard conception prevents from appreciating. If this venture does not, of course, aim to contest the results of algorithmic studies, it intends to enrich it with grounded sociological considerations.

笔记

Notes

1. 1.从我获得研究资助之日起，我的研究总体问题没有发生根本性的变化。

2. 1.  The general issue subtending my research has not fundamentally changed since the date at which I was awarded the research grant.

3. 2. CSF 的一个特点是其国际化视野。在我参加的官方活动中，院长们经常强调 CSF 吸引外国学生和研究人员的能力。在实验室中尤其如此，我是近一年来唯一的“本土”科学合作者。通用语言与这种国际环境相符；尽管实验室位于法语区，但大多数互动、演示和文件都是用英语进行的。

4. 2.  One of the particularities of the CSF was its international focus. During the official events I attended, deans regularly put forward the CSF’s capacity to attract foreign students and researchers. This was especially true in the case of the Lab where I was the only “indigenous” scientific collaborator for nearly a year. The lingua franca was in line with this international environment; even though the Lab was located in a French-speaking region, most interactions, presentations, and documents were in English.

5. 3.电荷耦合器件的发展历史已在Seitz 和 Einspruch (1998, 212–228) 和 Gertner (2013, 250–265) 中有所记载，尽管只是部分记载。

6. 3.  The history of the development of the charge-coupled device has been documented, though quite partially, in Seitz and Einspruch (1998, 212–228) and Gertner (2013, 250–265).

7. 4.有关 CCD 和图像传感器的简单介绍，请参阅 Allen 和 Triantaphillidou (2011, 155–173)。

8. 4.  For an accessible introduction to CCDs and image sensors, see Allen and Triantaphillidou (2011, 155–173).

9. 5. CMOS 是 CCD 的较新变体，其中每个像素包含一个光电探测器和一个放大器。此功能目前可以显著减小尺寸和功耗图像传感器。这也是为什么如今大多数便携式设备（如智能手机和小型相机）都配备 CMOS 的原因之一。

10. 5.  CMOS is a more recent variant of CCD where each pixel contains a photodetector and an amplifier. This feature currently allows significant size and power reduction of image sensors. This is one of the reasons why CMOSs now equip most portable devices such as smartphones and compact cameras.

11. 6.人们普遍认为，像素这个术语是“图片元素”的缩写，最早出现在 1969 年加州理工学院喷气推进实验室的一篇论文中（Leighton 等人，1969 年）。但实际上，这个术语在整个 20 世纪 60 年代经常在新兴的图像处理社区中使用，因此其历史要复杂得多。有关像素这个术语的简要历史，请参阅 Lyon (2006)。

12. 6.  It is commonly assumed that the term pixel, as a contraction of “picture element,” first appeared in a 1969 paper from Caltech’s Jet Propulsion Lab (Leighton et al. 1969). The story is more intricate than that as the term was regularly used in emergent image-processing communities thoughout the 1960s. For a brief history of the term pixel, see Lyon (2006).

13. 7.数字信号由n个维度表示，具体取决于用于描述信号的独立变量。例如，采样的数字声音通常被描述为一维信号，其因变量（振幅）随时间 ( t ) 而变化；数字图像通常被描述为二维信号，其因变量（强度）随两个轴 ( x , y ) 而变化，而视听内容将被描述为具有独立变量 ( x , y, t )的三维信号。有关数字信号处理的简单介绍，请参阅 Vetterli、Kovacevic 和 Goyal (2014)。

14. 7.  A digital signal is represented by n number of dimensions depending on the independent variables used to describe the signal. A sampled digital sound is, for example, typically described as a one-dimensional signal whose dependent variables—amplitudes—vary according to time (t); a digital image is typically described as a two-dimensional signal whose dependent variables—intensities—vary according to two axes (x, y), whereas audio-visual content will be described as a three-dimensional signal with independent variables (x, y, t). For an accessible introduction to digital signal processing, see Vetterli, Kovacevic, and Goyal (2014).

15. 8.这并非实验室的唯一研究重点。一些研究人员还致力于 CCD/CMOS 架构和传感器的研究。

16. 8.  It was not the only research focus of the Lab. Several researchers also worked on CCD/CMOS architectures and sensors.

17. 9。值得注意的是，要使数字图像处理和识别成为计算机科学的主要子领域，数字图像首先必须成为能够被计算机程序处理的稳定实体，这是一项长期的研究和开发工作。随着 CCD 和后来的 CMOS 等图像传感器的开发、标准化和工业生产，首先需要数据压缩方面的理论研究，例如 O'Neal Jr. (1966) 关于差分脉冲编码调制的研究、Ahmed、Natarajan 和 Rao (1974) 关于余弦变换的研究；或 Gray (1984) 关于矢量量化的研究。这些研究后来被纳入 1993 年批准的现已广泛使用的国际标准化组织规范 JPEG 的定义，这是另一个决定性的一步：从那一刻起，电信提供商、软件开发商和硬件制造商可以依赖和协调一种用于数字压缩静态图像表示的摄影编码技术（Hudson 等人，2017 年）。 20 世纪 90 年代末，微型计算机的普及、处理能力的逐步提升以及网络技术和标准的开发和维护也极大地促进了数字图像处理成为主流研究领域。因此，图像处理在研究、工业和国防领域的流行与多媒体通信设备的逐渐出现以及现在作为标准技术基础设施运行的基本组件的黑盒化有关。

18. 9.  It is important to note that for digital image processing and recognition to become a major subfield of computer science, digital images first had to become stable entities capable of being processed by computer programs—a long-standing research and development endeavor. Along with the development, standardization, and industrial production of image sensors such as CCDs and, later, CMOSs, theoretical works on data compression—such as those of O’Neal Jr. (1966) on differential pulse code modulation; Ahmed, Natarajan, and Rao (1974) on cosine transform; or Gray (1984) on vector quantization—have first been necessary. The later enrollment of these works for the definition of the now-widespread International Organization for Standardization norm JPEG, approved in 1993, was another decisive step: from that moment, telecommunication providers, software developers, and hardware manufacturers could rely on and coordinate around one single photographic coding technique for digitally compressed representations of still images (Hudson et al. 2017). During the late 1990s, the growing distribution of microcomputers, their gradual increase in terms of processing power, and the development and maintenance of web technologies and standards have also greatly contributed to establishing digital image processing as a mainstream field of study. The current popularity of image processing for research, industry, and defense is thus to be linked with the progressive advent of multimedia communication devices and the blackboxing of their fundamental components operating now as standard technological infrastructure.

19. 10.根据日本相机和影像产品协会（包括佳能、尼康、索尼和奥林巴斯等）的数据，数码相机的销量已从 2010 年的 6290 万台下降到不足2017 年为 2425 万张（Statista 2019）。然而，根据 InfoTrends 和 Bitkom 的估计，同期全球拍摄的照片数量从 6600 亿张增加到 12000 亿张（Richter 2017）。这种差异的原因之一是智能手机相机的日益复杂化以及 Instagram 和 Facebook 等社交媒体网站的普及和共享功能（Cakebread 2017）。

20. 10.  According to Japan-based industry association Camera & Imaging Products Association (to which, among others, Canon, Nikon, Sony, and Olympus belong), sales of digital cameras have dropped from 62.9 million in 2010 to fewer than 24.25 million in 2017 (Statista 2019). However, according to estimates generated by InfoTrends and Bitkom, the number of pictures taken worldwide increased from 660 billion to 1,200 billion over the same period (Richter 2017). This discrepancy is due, among other things, to the increasing sophistication of smartphone cameras as well as the popularity and sharing functionalities of social-media sites such as Instagram and Facebook (Cakebread 2017).

21. 11.例如，Google、Amazon、Apple、Microsoft和IBM均提出了用于图像识别的应用程序编程接口产品（分别是Cloud Vision、Amazon Rekognition、Apple Vision、Microsoft Computer Vision和Watson Visual Recognition）。

22. 11.  For example, Google, Amazon, Apple, Microsoft, and IBM all propose application programming interface products for image recognition (respectively, Cloud Vision, Amazon Rekognition, Apple Vision, Microsoft Computer Vision, and Watson Visual Recognition).

23. 12.根据爱德华·斯诺登 (Edward Snowden) 获得的 2011 年文件，美国国家安全局在 2010 年每天截取数百万张图像，以开发针对恐怖嫌疑人的计算机化追踪方法 (Risen and Poitras 2014)。中国当局还大量投资于面部识别技术，用于安全和控制目的 (Mozur 2018)。

24. 12.  According to 2011 documents obtained by Edward Snowden, the National Security Agency intercepted millions of images per day throughout the year 2010 to develop computerized tracking methods for suspected terrorists (Risen and Poitras 2014). Chinese authorities also heavily invest in facial recognition for security and control purposes (Mozur 2018).

25. 13.例如，请参阅《国际计算机视觉杂志》、《IEEE 模式分析与机器智能学报》、《IEEE 图像处理学报》或《模式识别》。

26. 13.  See, for example, International Journal of Computer Vision, IEEE Transactions on Pattern Analysis and Machine Intelligence, IEEE Transactions on Image Processing, or Pattern Recognition.

27. 14.例如，请参阅IEEE 计算机视觉与模式识别会议、欧洲计算机视觉会议、IEEE 国际计算机视觉会议或 IEEE 国际图像处理会议。

28. 14.  See, for example, IEEE Conference on Computer Vision and Pattern Recognition, European Conference on Computer Vision, IEEE International Conference on Computer Vision, or IEEE International Conference on Image Processing.

29. 15.以图像处理算法方面学术界和工业界之间的密切关系为例，专门从事自动结账图像识别的初创公司 Standard Cognition 的首席执行官乔丹·费舍尔 (Jordan Fisher) 在最近的TechCrunch文章 (Constine 2019) 中说道：“这是狂野的西部——应用刚刚发布的尖端、最先进的机器学习研究。我们阅读论文，然后在论文发表几周后实施，将这些想法付诸实践，使其具有生产价值。”

30. 15.  Giving an example of the close relationships between academic and industrial worlds regarding image-processing algorithms, Jordan Fisher—chief executive officer of Standard Cognition, a start-up that specializes in image recognition for autonomous checkout—says in a recent TechCrunch article (Constine 2019): “It’s the wild west—applying cutting-edge, state-of-the-art machine learning research that’s hot off the press. We read papers then implement it weeks after it’s published, putting the ideas out into the wild and making them production-worthy.”

31. 16. 2016年和2017年，苹果和微软研究团队的论文均获得了图像处理和识别领域最负盛名的会议IEEE计算机视觉与模式识别会议的最佳论文奖。此外，2018年，谷歌推出了自己的学术期刊《Distill Research Journal》，旨在推动图像和视频识别领域的机器学习。

32. 16.  In 2016 and 2017, papers from Apple and Microsoft research teams won the best-paper award of the IEEE Conference on Computer Vision and Pattern Recognition, the most prestigious conference in image processing and recognition. Moreover, in 2018, Google launched Distill Research Journal, its own academic journal aiming at promoting machine learning in the field of image and video recognition.

33. 17.例如，在 Knuth (1997a) 中，作者首先回顾说，“算法”是“算法主义”一词的后期演变，而“算法主义”本身源于著名波斯数学家 Ab ū 'Abd All ā h Muhammad ibn M ū sa al-Khw ā rizmi 的名字——字面意思是“阿卜杜拉之父，穆罕默德，摩西之子，Khw ā rizm人”，Khw ā rizm 在本例中指的是咸海以南的一个地区 (Zemanek 1981)。Knuth 随后指出，从其最初的含义来看，随着使用阿拉伯数字进行算术运算的过程，算法（algorism）一词逐渐被讹传：“正如牛津英语词典所解释的，该词‘经历了许多伪词源学曲解，包括最近的算法，其中它被学识性地与算术（arithmetic）一词的希腊词根混淆了’ ” (Knuth 1997a, 2)。

34. 17.  This is for example the case in Knuth (1997a) where the author starts by recalling that “algorithm” is a late transformation of the term “algorism” that itself derives from the name of famous Persian mathematician Abū ‘Abd Allāh Muhammad ibn Mūsa al-Khwārizmi—literally, “Father of Abdullah, Mohammed, son of Moses, native of Khwārizm,” Khwārizm referring in this case to a region south of the Aral Sea (Zemanek 1981). Knuth then specifies that from its initial acceptation as the process of doing arithmetic with Arabic numerals, the term algorism gradually became corrupted: “as explained by the Oxford English Dictionary, the word ‘passed through many pseudo-etymological perversions, including a recent algorithm, in which it is learnedly confused’ with the Greek root of the word arithmetic” (Knuth 1997a, 2).

35. 18.例如，参见 Knuth (1997, 4) 对算法的 (非常) 临时定义：“算法的现代含义与配方、过程、方法、技术、程序、例行程序、繁文缛节非常相似。 ”

36. 18.  See, for example, the (very) temporary definition of algorithms by Knuth (1997, 4): “The modern meaning for algorithm is quite similar to that of recipe, process, method, technique, procedure, routine, rigmarole.”

37. 19.例如，请参阅 Sedgewick 和 Wade (2011, 3) 对算法的定义：“适合计算机实现的解决问题的方法”。

38. 19.  See, for example, Sedgewick and Wade’s (2011, 3) definition of algorithms as “methods for solving problems that are suited for computer implementation.”

39. 20.另请参阅 Cormen 等人（2009，5）的定义：“一个定义明确的计算过程，它将某个值或一组值作为输入，并产生某个值或一组值作为输出，因此是将输入转换为输出的一系列计算步骤。”

40. 20.  See also Cormen et al.’s (2009, 5) definition: “A well-defined computational procedure that takes some value, or set of values, as input and produces some value, or set of values, as output [being] thus a sequence of computational steps that transform the input into the output.”

41. 21 . 另见 Dasgupta、Papadimitriou 和 Vazirani (2006, 12) 的表述：“每当我们有一个算法时，我们总是会问三个问题：1. 它正确吗？2. 作为 n 的函数，它需要多少时间？ 3. 我们能做得更好吗？” 以及 Skiena (2008, 4)：“一个好的算法有三个理想的特性。我们寻求正确、高效且易于实现的算法。”

42. 21.  See also Dasgupta, Papadimitriou, and Vazirani’s (2006, 12) phrasing: “Whenever we have an algorithm, there are three questions we always ask about it: 1. Is it correct? 2. How much time does it take, as a function of n? 3. And can we do better?” And also Skiena (2008, 4): “There are three desirable properties for a good algorithm. We seek algorithms that are correct and efficient, while being easy to implement.”

进入实验室餐厅

Entering the Lab’s Cafeteria

2013 年 11 月 7 日下午 3 点左右，我（FJ）走进实验室餐厅参加第一次小组会议。那时，小组和项目主题已经确定：实验室同事之间的先前讨论一致认为，关于显著性检测的新集体出版物与 CL、GY 和 BJ 的最新技术以及专业知识相关。自然，就像任何刚到现场的民族志学者一样，我非常焦虑：我能否达到期望？他们会帮助我理解他们所做的事情吗？我参与该项目显然是一个自上而下的决定，因为实验室主任已将我分配到该项目以帮助我正确开始我的调查。小组会欢迎我吗？我试图阅读 CL 之前发给我的一些关于显著性检测的论文，但我被他们心照不宣的假设搞糊涂了。既然数字图像中重要的事物因人而异，那么如何才能检测到这种称为“显著性”的奇怪事物呢？论文算法似乎依赖的“基本事实”这个奇怪的概念又是什么呢？“基本事实”和“事实”：对于 STS 学者来说，这样的结合听起来非常成问题。然而，我一走进实验室的自助餐厅，小组成员就向我介绍了该项目的雄心以及他们打算如何运行它：²

Around 3 p.m. on November 7, 2013, I (FJ) entered the Lab’s cafeteria for the first Group meeting. By that time, the Group and the topic of the project had already been defined: previous discussions among the Lab associates agreed that a new collective publication in saliency detection was relevant regarding the state of the art as well as the expertise of CL, GY, and BJ. Naturally, as any ethnographer freshly landed on his field site, I was terribly anxious: Would I live up to the expectations? Would they help me understand what they do? My participation in the project was clearly a top-down decision as the Lab’s director had assigned me to the project to help me properly start my inquiry. Would the Group welcome me? I tried to read some papers on saliency detection that CL previously sent me but I was confused by their tacit postulates. How would it be possible to detect this strange thing called “saliency” since what is important in a digital image certainly varies from person to person? And what is this odd notion of “ground truth” that the papers’ algorithms seem to rely on? “Ground” and “truth”: for an STS scholar, such a conjunction sounded highly problematic. As soon as I entered the Lab’s cafeteria though, the members of the Group presented me with the ambitions of the project and how they intended to run it:²

小组会议，实验室餐厅，2013 年 11 月 7 日

Group meeting, the Lab’s cafeteria, November 7, 2013

氯：  “那么你听说过显著性，对吧？”

CL:  “So you heard about saliency, right?”

缩略词：  “嗯，我读了一些内容。”

FJ:  “Well, I’ve read some stuff.”

氯：  “这是一个很大的话题，但基本上，当你看一张图片时，通常并不是所有的东西都很重要，你只会关注一些元素。……我们试图做的基本上就像一个模型，它可以检测图像中应该引起注意的元素。…… GY 研究了一种使用对比来分割物体的模型，BJ 有一个检测人脸的模型。我们将以它们为基础。……目前，大多数显著性模型只检测物体，而不关注人脸。这没有基本事实。但我们说的是，人脸也很重要，通常会直接吸引注意力。……这就是重点：我们基本上想把人脸纳入显著性中。”

CL:  “Huge topic, but basically, when you look at an image, not everything is important usually, and you focus only on some elements. … What we try to do basically, it’s like a model that detects elements in an image that should attract attention. … GY’s worked on a model that uses contrasts to segment objects and BJ has a model that detects faces. We’ll use them as a base. … For now, most saliency models only detect objects and don’t pay attention to faces. There’s no ground truth for that. But what we say is that faces are also important and usually attract directly the attention. … And that’s the point: we want to include faces to saliency, basically.”

戈瑞：  “并且分割人脸。因为人脸检测器只输出矩形。......该模型可以有很多应用，例如显示或压缩。”

GY:  “And segment faces. Because face detectors output only rectangles. … There can be many applications [for the model], like in display or compression for example.”

许多问题立即浮现出来。为什么以及为什么关注“应该引起注意的元素”很重要？为什么没有“基本事实”来检测“多个物体和面孔”会成问题？基本事实到底是什么？为什么它与“显著性”及其潜在的工业应用有关？在调查的早期阶段，民族志的曲折流动在某种程度上剥夺了我们的地标。为了跟随小组并能够充分探索这些材料，显然需要更多的设备。因此，我将暂时“暂停”小组项目的叙述，并考虑一下小组项目框架所依据的显著性检测的社会历史背景。一旦掌握了这些介绍性要素，我就会回到第一次小组会议。

Many questions immediately arose. How and why is it important to focus on “elements that should attract attention”? Why is it problematic not to have a “ground truth” to detect “multiple objects and faces”? And what is a ground truth anyway? Why is it related to “saliency” and its potential industrial applications? Already at this early stage of the inquiry, the meandering flows of ethnography somewhat deprive us from our landmarks. To follow the Group and become able to fully explore these materials, some more equipment is obviously needed. I will thus temporally “pause” the account of the Group’s project and consider for a while the sociohistorical background of saliency detection that underlies the Group’s framing of its project. Once these introductory elements are acquired, I will be come back to this first Group meeting.

幕后元素：显著性检测和数字图像处理

Backstage Elements: Saliency Detection and Digital Image Processing

对于从事图像处理的计算机科学家来说，“显著性”是一个模糊的术语，其历史难以追溯。“显著性”一词是通过跨越不同但密切相关的研究领域而逐渐创造出来的。一个出发点可能是 20 世纪 70 年代，当时在认知心理学和神经生物学³中发展起来的解释模型开始图式化人类大脑如何快速处理远远超出其估计处理能力的视觉数据量（Eason、Harter 和 White 1969；Lappin 和 Uttal 1976；Shiffrin 和 Gardner 1972）。4^经过许多争论和争议，关于人类“选择性视觉注意方法”的整体过程逐渐达成了粗略的共识，区分了选择和门控视觉信息的两个神经过程（Itti 和 Koch 2001；Heinke 和 Humphreys 2004）。⁵一方面，存在与任务无关的、快速的“自下而上的视觉注意过程”，该过程选择显着的刺激，例如颜色对比、特征方向或空间频率。另一方面，存在较慢的“自上而下的视觉注意过程”，该过程根据要完成的任务选择性地运行。“显著性图”一词由 Koch 和 Ullman (1985) 提出，用于定义大脑自下而上的视觉注意过程的最终结果。

“Saliency” for computer scientists in image processing is a blurry term with a history that is difficult to track. The term “saliency” was gradually created by straddling different—yet closely related—research areas. One point of departure could be the 1970s when explicative models developed in cognitive psychology and neurobiology³ started to schematize how the human brain could quickly handle an amount of visual data that is far larger than its estimated processing capabilities (Eason, Harter, and White 1969; Lappin and Uttal 1976; Shiffrin and Gardner 1972).⁴ After many disputes and controversies, a rough agreement about the overall process of humans’ “selective visual attention method” had progressively emerged that distinguishes between two neuronal processes of selecting and gating visual information (Itti and Koch 2001; Heinke and Humphreys 2004).⁵ On the one hand, there is a task-independent and rapid “bottom-up visual attention process” that selects conspicuous stimuli such as color contrasts, feature orientations, or spatial frequency. On the other hand, there is a slower “top-down visual attention process” that operates selectively based on tasks to accomplish. The term “saliency map” was proposed by Koch and Ullman (1985) to define the final result of the brain’s bottom-up visual attention process.

20 世纪 80 年代，认知心理学家和神经生物学家提出了大脑处理光信号的两种不同“路径”理论——一种是快速而通用的，另一种是较慢且针对特定任务的——这启发了科学家，他们的机器面临着计算机视觉方面的类似问题：从 CCD 发出的采样数字信号流太大，无法一次性处理。从这一点开始，两类不同的图像处理检测算法逐渐成型。第一类算法受到假设的自下而上的视觉注意示意图过程的启发，并试图检测给定图像像素中刻画的“低级特征”，例如强度、颜色、方向和纹理。⁶经过 21 世纪 Laurent Itti 和 Christof Koch 的学术努力（Itti、Koch 和 Niebur 1998；Itti、Koch 和 Braun 2000；Itti 和 Koch 2001；Elazary 和 Itti 2008；Zhao 和 Koch 2011），“显著性”一词逐渐融入了第一类算法，并被称为显著性检测算法。第二类图像处理检测算法受到假设的自上而下的视觉注意图式过程的启发，并基于“高级特征”，这些特征必须由机器根据以下规则进行学习：特定指标（例如，人脸或汽车检测）。这通常涉及自动学习程序和日益庞大的数据库的管理（Grimson 和 Lozano-Perez 1983；Lowe 1999）。

In the 1980s, the way that cognitive psychologists and neurobiologists theorized two different “paths” for the brain to process light signals—one fast and generic, the other slower and task-specific—inspired scientists whose machines face a similar problem in computer vision: the stream of sampled digital signals that emanated from CCDs were too large to be processed all at once. From this point, two different classes of image-processing detection algorithms have progressively been shaped. The first class was inspired by the assumed bottom-up schematic process of visual attention and tried to detect “low-level features” inscribed within the pixels of a given image, such as intensity, color, orientation, and texture.⁶ Through the academic efforts of Laurent Itti and Christof Koch in the 2000s (Itti, Koch, and Niebur 1998; Itti, Koch, and Braun 2000; Itti and Koch 2001; Elazary and Itti 2008; Zhao and Koch 2011), the term “saliency” was progressively assimilated into this first class of algorithms that became labeled saliency-detection algorithms. The second class of image-processing detection algorithms was inspired by the assumed top-down schematic process of visual attention and is based on “high-level features” that have to be learned by machines according to specific metrics (e.g., face or car detection). This often involves automated learning procedures and the management of increasingly large databases (Grimson and Lozano-Perez 1983; Lowe 1999).

尽管在底层结构方面存在差异，高级和低级检测算法都与相同的构建工作流程相关，该工作流程由五个相互关联且存在问题的步骤组成：

Despite differences in terms of substratum, both high-level and low-level detection algorithms were, and are, bound to the same construction workflow that consists of five interrelated and problematic steps:

1. 获取有限数据集。
1. The acquisition of a finite dataset.
2. 在此数据集的数据上，手动标记明确的目标，这里将其定义为所需算法将被要求检测的元素（面部，汽车，显著区域）。
3. On the data of this dataset, the manual labeling of clear targets, defined here as the elements (faces, cars, salient regions) the desired algorithm will be asked to detect.
4. 构建一个数据库，收集未标记数据及其手动标记的对应数据。研究界通常将该数据库称为“ground truth”。
5. The construction of a database gathering the unlabeled data and their manually labeled counterparts. This database is usually called “ground truth” by the research community.
6. 基于真实数据库的代表性部分设计算法的计算性质和参数。
7. The design of the algorithm’s calculating properties and parameters based on a representative part of the ground-truth database.
8. 根据其余真实数据库对算法的性能进行评估。
9. The evaluation of the algorithm’s performances based on the rest of the ground-truth database.

为了说明这一示意性工作流程，我们假设存在φ，这是图像处理中的一种标准检测算法。φ 的存在取决于一组有限的数字图像，人类工作者之前已为这些图像标记了目标（例如，人脸、汽车、显著区域）。然后将未标记的图像及其手动标记的对应图像收集到数据库中，以形成φ的地面实况。为了设计和编码φ，地面实况被随机分成两部分：“训练集”和“评估集”。φ 的设计者将使用训练集提取有关目标的形式信息，通常借助数学表达式。在制定并翻译成机器可读的代码后，算法φ将在评估集上进行测试，以查看它检测未用于设计其属性的目标的效果如何。通过与评估集的对比，由于之前的人工标记工作，φ产生了精确数量的输出，这些输出可以归类为“真阳性”、“假阴性”或“误报”。通过对手动设计的目标和自动产生的输出进行比较，可以获得诸如精度（先前定义为目标的检测到的项目的比例）和召回率（检测到的项目中所占目标的比例）等统计指标，以比较和排名竞争算法（见图2.1）。

To illustrate this schematic workflow, let us hypothesize the existence of ϕ, a standard detection algorithm in image processing. The very existence of ϕ depends upon a finite set of digital images for which human workers have previously labeled targets (e.g., faces, cars, salient regions). The unlabeled images and their manually labeled counterparts are then gathered together within a database to form the ground truth of ϕ. To design and code ϕ, the ground truth is randomly split into two parts: the “training set” and the “evaluation set.” The designers of ϕ would use the training set to extract formal information about the targets, often with help of mathematical expressions. Once formulated and translated into machine-readable code, the algorithm ϕ is tested on the evaluation set to see how well it detects targets that were not used to design its properties. From its confrontation with the evaluation set, ϕ produces a precise number of outputs that can be qualified either as “true positives,” “false negatives,” or “false positives,” thanks to the previous human-labeling work. Out of this comparison between manually designed targets and automatically produced outputs, statistical measures such as precision (the fraction of detected items that were previously defined as targets) and recall (the fraction of targets among the detected items) can be obtained to compare and rank competing algorithms (see figure 2.1).

高级检测算法的一个缺点是它们是针对特定任务的，无法单独检测不同类型的目标：人脸检测算法可以检测人脸，汽车检测算法可以检测汽车，飞机检测算法可以检测飞机，等等。7^然而，这种高级检测算法的一个好处是，对于设计它们的人来说，它们的目标（人脸、汽车、飞机）的定义通常涉及轻微的歧义：汽车、人脸或飞机具有相当明确的特征，有助于达成一致。然后，计算机科学家可以手动塑造目标和基本事实，以训练高级检测算法。此外，这些基本事实还可以作为竞争高级检测算法的裁判，因为它们提供了精确度和召回率指标。人脸检测子领域拥有众多的基本事实和算法命题，为高度复杂的算法提供了一个典型的例子。自 2000 年代以来，图像处理领域就已经发展并形成了具有竞争力的课题（见图2.2）。

One drawback of high-level detection algorithms is that they are task-specific and cannot by themselves detect different types of targets: a face-detection algorithm will detect faces, a car-detection algorithm will detect cars, a plane-detection algorithm will detect planes, and so on.⁷ Yet, one of the benefits of such high-level detection algorithms is that the definition of their targets (faces, cars, planes) often involves minor ambiguities for those who design them: cars, faces, or planes have rather unambiguous characteristics that facilitate agreement. Targets and ground truths can then be manually shaped by computer scientists in order to train high-level detection algorithms. Moreover, these ground truths can also serve as referees among competing high-level detection algorithms as they provide precision and recall metrics. The subfield of face detection with its numerous ground truths and algorithmic propositions provides a paradigmatic example of a highly developed and competitive topic in image processing since at least the 2000s (see figure 2.2).

在 21 世纪，与高级检测研究不同，低级显著性检测没有“自然”的基本事实，无法设计和评估计算模型。8^当时，如果显著性检测的任务独立性和自适应性在理论上对自动图像裁剪（Santella 等人，2006 年）、小型设备上的自适应显示（Chen 等人，2003 年）、广告设计和图像压缩（Itti，2000 年）很有吸引力，那么由于缺乏可以训练和评估计算模型的基本事实，显著性检测无法成为数字图像处理中的活跃主题。正如 Itti、Koch 和 Niebur（1998 年）在自然图像上测试第一个显著性检测算法时所承认的那样：

In the 2000s, unlike research in high-level detection, low-level saliency detection had no “natural” ground truth allowing the design and evaluation of computational models.⁸ At that time, if the task-independent and adaptive character of saliency detection was theoretically interesting for automatic image cropping (Santella et al. 2006), adaptive display on small devices (Chen et al. 2003), advertising design, and image compression (Itti 2000), the absence of any ground truth that could allow the training and evaluation of computational models prevented saliency detection from being an active topic in digital image processing. As Itti, Koch, and Niebur (1998) confessed when they tested the very first saliency-detection algorithm on natural images:

对于许多这样的[自然]图像，很难客观地评估模型，因为没有可供比较的客观参考，并且观察者可能对哪些位置最突出存在分歧。（Itti、Koch 和 Niebur 1998，1258；斜体添加）

With many such [natural] images, it is difficult to objectively evaluate the model, because no objective reference is available for comparison, and observers may disagree on which locations are the most salient. (Itti, Koch, and Niebur 1998, 1258; italics added)

自然图像中的显著性检测是一个模棱两可的话题，很难用基本事实来表达。虽然定义用于训练和评估高级人脸检测或汽车检测算法的明确目标通常很简单（但也很耗时），但对于显著性检测算法来说，这样做要复杂得多，因为自然图像中被视为显著的东西往往因人而异。虽然在 2000 年代，显著性检测算法可能对许多工业应用很有前景，但图像处理领域没有人找到为自然图像设计基本事实的方法。

Saliency detection in natural images is an equivocal topic not easily expressed in a ground truth. Whereas it is usually straightforward (and yet time consuming) to define univocal targets for training and evaluating high-level face-detection or car-detection algorithms, it is far more complex to do so for saliency-detection algorithms because what is considered as salient in a natural image tends to change from person to person. While in the 2000s saliency-detection algorithms might have been promising for many industrial applications, no one in the field of image processing had found a way to design a ground truth for natural images.

2007 年，刘等人针对这一问题提出了一种创新解决方案，并创建了自然图像中显著性检测的第一个基本事实。他们的转变很聪明，代价高昂，并为图像处理文献中显著性检测子领域的构建和建立做出了巨大贡献。刘等人的第一步是通过结合高级检测的概念，提出了显著性检测的一种可能范围。据他们介绍，显著性的计算模型不是试图突出数字图像中的显著区域，而是应该检测给定数字图像中最显著的对象。因此，他们将显著性问题定义为二元和一次性对象相关。据他们介绍，为了解决显著性检测的僵局，显著性检测算法应该将一个显著对象与图像的其余部分区分开来：

In 2007, Liu et al. proposed an innovative solution to this issue and created the very first ground truth for saliency detection in natural images. Their shift was smart, costly, and contributed greatly to framing and establishing the subfield of saliency detection in the image-processing literature. Liu et al.’s first move was to propose one possible scope of saliency detection by incorporating concepts from high-level detection. According to them, instead of trying to highlight salient areas within digital images, computational models for saliency should detect the most salient object within a given digital image. They thus framed the saliency problem as being binary and one-off object related. According to them, to get around the impasse of saliency detection, saliency-detection algorithms should distinguish one salient object from the rest of the image:

我们将显著物体的高级概念融入到每幅图像的视觉注意过程中。我们称它们为显著物体，或我们熟悉的前景物体。……我们将显著物体检测公式化为二元标记问题，将显著物体与背景区分开来。与人脸检测一样，我们检测熟悉的物体；与人脸检测不同的是，我们在图像中检测熟悉但未知的物体。（刘等人，2007 年，1-2）

We incorporate the high-level concept of salient object into the process of visual attention in each respective image. We call them salient objects, or foreground objects that we are familiar with. … We formulate salient object detection as a binary labelling problem that separates a salient object from the background. Like face detection, we detect a familiar object; unlike face detection, we detect a familiar yet unknown object in an image. (Liu et al. 2007, 1–2)

由于对显著性概念的这种改进（从“任何首先引起注意的事物”到“图片中首先引起注意的物体”），刘等人可以组织一项实验，以构建由计算模型检索的合法目标。他们首先从互联网论坛和搜索引擎中随机收集了 130,099 张高质量自然图像。然后他们手动选择了 20,840 张符合他们对显著性问题的定义：根据他们的说法，图像只包含一个显著对象。这个初始选择操作至关重要，因为它排除了具有多个潜在显著对象的图像。结果是初始数据集没有具有混合特征的复杂图片（见图2.3）。

Thanks to this refinement of the concept of saliency (from “anything that first attracts attention” to “the one object in a picture that first attracts attention”), Liu et al. could organize an experiment in order to construct legitimate targets to be retrieved by computational models. They first randomly collected 130,099 high-quality natural images from internet forums and search engines. Then they manually selected 20,840 images that fit with their definition of the saliency problem: images that, according to them, contained only one salient object. This initial selection operation was crucial as it excluded images with several potential salient objects. The result was an initial dataset of no complex pictures with mixed features (see figure 2.3).

然后他们分两步进行。首先，他们请三名工作人员在他们认为是每幅图像中最显著的物体上手动绘制一个矩形。对于每幅图像，刘等人随后获得三个不同的矩形，其一致性可以通过共享像素的百分比来衡量。对于给定图像，如果其三个矩形比所选阈值更一致（这里是 80% 的像素相同），则该图像被视为包含“高度一致的显著物体”（刘等人，2007，2）。经过这个第一个选择步骤后，他们的数据集（称为α）包含大约一万三千张图像。

They then proceeded in two steps. First, they asked three human workers to manually draw a rectangle on what they thought was the most salient object in each image. For each image, Liu et al. then obtained three different rectangles whose consistencies could be measured by the percentage of shared pixels. For a given image, if its three rectangles were more consistent than a chosen threshold (here, 80 percent of pixels in common), the image was considered as containing a “highly consistent salient object” (Liu et al. 2007, 2). After this first selection step, their dataset called α contained around thirteen thousand images.

第二步，刘等人从α中随机选择了五千张高度一致的显著物体图像，创建了第二个数据集β 。然后，他们要求另外九名人类工作者用一个矩形标记β中每幅图像的显著物体。这一次，刘等人为每幅图像获得了九个不同但高度一致的矩形，这些矩形的平均表面被视为它们的“显著性概率图”（刘等人，2007，3）。得益于这种构建的社会共识，五千张显著性概率图（从计算机科学的角度来看，是由特定数值构成的有形矩阵）可以被视为他们所构建的显著性问题的最佳解决方案。整个基本事实（收集自然图像及其对应图像的数据库）显著性概率图成为算法得以实现的物质基础，通过构建这一基本事实，刘等人定义了一个新问题的术语，该问题的解可以通过计算方法得到。

For the second step, Liu et al. randomly selected five thousand highly consistent salient-object images from α to create a second dataset called β. They then asked nine other human workers to label the salient object of every image in β with a rectangle. This time, Liu et al. obtained for every image nine different yet highly consistent rectangles whose average surface was considered their “saliency probability map” (Liu et al. 2007, 3). Thanks to this constructed social agreement, the five thousand saliency probability maps—in a computer science perspective, tangible matrices constituted of specific numerical values—could then be considered the best solutions to the saliency problem as they framed it. The whole ground truth—the database gathering the natural images and their corresponding saliency probability maps—became the material base on which the desired algorithm could be developed. By constructing this ground truth, Liu et al. defined the terms of a new problem whose solutions could be retrieved by means of calculating methods.

这里的转变并非微不足道。事实上，通过组织这次调查，邀请人们到他们的实验室，欢迎他们，向他们解释主题，编写适当的计算机程序让他们标记图像，并将结果收集到适当的数据库中以便进行统计处理，刘等人将他们最初简化的显著性检测概念转变为具有具体数值的可行且明确的目标。在这个费力的过程结束时，刘等人可以从集合α中随机选择两千张图像，从集合β中随机选择一千张图像来构建一个训练集（刘等人，2007，5-6），以分析他们构建的但因一致而合理的目标的共同特征。一旦从训练集的目标中提取出足够的数值特征并以机器可读的语言实现，他们就会使用集合β中剩余的四千张图像来统计测量其算法的性能。此外，他们还首次将其算法的检测性能与其他实验室已经提出的两种竞争算法进行比较，但由于缺乏与显著性相关的任何“自然”目标，这些算法之前无法在自然图像上进行评估。除了实际完成显著性检测算法之外，刘等人的重大创新在于重新定义了显著性问题，以便可以进行性能评估（见图2.4）。

The shift here was not trivial. Indeed, by organizing this survey, inviting people into their laboratory, welcoming them, explaining the topic to them, writing the appropriate computer programs to make them label the images, and gathering the results in a proper database in order to statistically process them, Liu et al. transformed their initial reduced conception of saliency detection into workable and unambiguous targets with specific numerical values. At the end of this laborious process, Liu et al. could randomly select two thousand images from set α and one thousand images from set β to construct a training set (Liu et al. 2007, 5–6) to analyze the shared features of their constructed-yet-sound-by-virtue-of-agreement targets. Once the adequate numerical features were extracted from the targets of the training set and implemented in machine-readable language, they used the four thousand remaining images from set β to statistically measure the performances of their algorithm. Further, and for the very first time, they also could compare the detection performances of their algorithm with two competing algorithms that had already been proposed by other laboratories but that could not have been evaluated on natural images before due to the lack of any “natural” targets related to saliency. Besides the actual completion of their saliency-detection algorithm, the great innovation of Liu et al. was then to redefine the saliency problem so that it could allow performance evaluations (see figure 2.4).

通过发表论文并在网上公开提供他们的理论基础，毫不夸张地说，刘等人在图像处理领域建立了一个新的可评估的研究方向。他们已经建立了一个昂贵的基础设施，随时可以重新用于支持其他竞争算法命题，根据刘等人的理论基础和它所包含的显著性定义，这些算法命题的性能可能更好。他们的出版物不仅仅是一篇论文：这是一篇允许其他论文发表的论文，因为它们提供了一个理论基础，其他研究人员可以使用它，只要他们正确引用这篇开创性的论文并接受理论基础的受限但可操作的显著性定义^。9

By publishing their paper and also publicly providing their ground truth online, it is not an exaggeration to say that Liu et al. established a newly assessable research direction in image processing. A costly infrastructure had been put together, ready to be reused to support other competing algorithmic propositions with perhaps better performances according to Liu et al’s ground truth and the definition of saliency it encapsulates. Their publication was more than a paper: it was a paper that allowed other papers to be published as they provided a ground truth that could be used by other researchers as long as they properly quote the seminal paper and accept the ground truth’s restricted—yet operational—definition of saliency.⁹

另一篇关于显著性检测的重要论文（因此也是我们很快将继续关注的小组项目）已发表2008 年，Wang 和 Li 提出了显著性检测新方法。在他们看来，尽管 Liu 等人 (2007) 将显著性问题定义为二元问题的做法是正确的，但他们的边界框基本事实仍然不能令人满意，因为它很可能会评估不准确的结果（见图2.5）。为了改进 Liu 等人的第一个显著性检测基本事实的测量方法，Wang 和 Li 从β数据集中随机选择了 300 张图像，并使用分割工具手动标记这 300 个显著物体的轮廓。他们随后提出并评估了一种显著性检测算法，该算法“不仅能捕捉显著物体的粗略位置和区域，而且还能粗略地保持轮廓正确”（Wang 和 Li 2008，965）。

Another important paper for saliency detection—and therefore also for the Group’s project that we shall soon continue to follow—was published in 2008 by Wang and Li. To them, even though Liu et al. (2007) were right to frame the saliency problem as a binary problem, their bounding-box ground truth remained unsatisfactory as it could well evaluate inaccurate results (see figure 2.5). To refine the measures of Liu et al.’s very first ground truth for saliency detection, Wang and Li randomly selected three hundred images from β dataset and used a segmentation tool to manually label the contours of each of the three hundred salient objects. What they proposed and evaluated then was a saliency-detection algorithm that “not only captures the rough location and region of the salient objects, but also roughly keeps the contours right” (Wang and Li 2008, 965).

从这一点来看，图像处理中的显著性检测几乎已经确定：尽管后来提出了许多利用不同低级像素信息的算法（Achanta 等人 2009；Chang 等人 2011；Cheng 等人 2011；Goferman、Zelnik-Manor 和 Tal 2012；Shen 和 Wu 2012；Wang 等人 2010），但它们都与刘等人在 2007 年定义的显著性问题息息相关。尽管后来在发表的论文中提出了其他基本事实（Judd、Durand 和 Torralba 2012；Movahedi 和 Elder 2010）以拓宽显著性检测的范围（特别是通过提出两个可以偏心的对象的图像），但刘等人将显著性检测作为二元对象相关问题的开创性框架仍然没有受到挑战。当该集团于 2013 年 11 月启动他们的项目时，刘等人对显著性问题的难题化继续支持了在速度和准确性方面有所差异的算法之间的竞争（见图2.6）。

From this point, saliency detection in image-processing was almost set: even though many algorithms exploiting different low-level pixel information were later proposed (Achanta et al. 2009; Chang et al. 2011; Cheng et al. 2011; Goferman, Zelnik-Manor, and Tal 2012; Shen and Wu 2012; Wang et al. 2010), they were all bound to the saliency problem as defined by Liu et al. in 2007. And even though other ground truths have later been proposed in published papers (Judd, Durand, and Torralba 2012; Movahedi and Elder 2010) to widen the scope of saliency detection (notably by proposing images with two objects that could be decentered), Liu et al.’s seminal framing of saliency detection as a binary object-related problem remained unchallenged. And when the Group started their project in November 2013, Liu et al.’s problematization of the saliency problem was continuing to support a competition among algorithms that differentiated themselves by speed and accuracy (see figure 2.6).

了解了图像处理中显著性的简要历史，我们就可以更好地跟踪该小组，因为它试图构建自己的创新显著性检测算法。社会调查、轮廓清晰的显著物体定义竞争算法的目标、与显著性的二元问题化相关的基本事实、有前景的工业应用：我们即将探索的阶段得到了所有这些要素的支持，从而约束了小组成员在项目的塑造过程中，并为他们提供进一步重新配置的机会。

With this brief history of saliency in image processing, we are better equipped to follow the Group as it tries to construct its own innovative saliency-detection algorithm. Social surveys, salient objects whose contours define the targets of competing algorithms, ground truths bound to a binary problematization of saliency, promising industrial applications: the stage we are about to explore is supported by all of these elements, constraining the members of the Group in the shaping of their project as well as providing them opportunities for further reconfigurations.

重新定义显著性

Reframing Saliency

如果在本章开头，该小组的解释显得相当隐晦，那么之前的介绍性回顾现在应该能够让我们批判性地理解它们。让我们再看一遍同一段摘录：

If, at the beginning of the chapter, the Group’s explanations appeared quite cryptic, the previous introductory review should now enable us to understand them critically. Let us thus look at the same excerpt once again:

小组会议，实验室餐厅，2013 年 11 月 7 日

Group meeting, the Lab’s cafeteria, November 7, 2013

氯：  “那么，你听说过显著性，对吧？”

CL:  “So, you heard about saliency, right?”

缩略词：  “嗯，我读了一些内容。”

FJ:  “Well, I’ve read some stuff.”

氯：  “这是一个很大的话题，但基本上，当你看一张图片时，通常并不是所有的东西都很重要，你只会关注一些元素。……我们试图做的基本上就像一个模型，它可以检测图像中应该引起注意的元素。…… GY 研究了一种使用对比来分割物体的模型，BJ 有一个检测人脸的模型。我们将以它们为基础。……目前，大多数显著性模型只检测物体，而不关注人脸。这没有基本事实。但我们说的是，人脸也很重要，通常会直接吸引注意力。……这就是重点：我们基本上想把人脸纳入显著性中。”

CL:  “Huge topic, but basically, when you look at an image, not everything is important usually, and you focus only on some elements. … What we try to do basically, it’s like a model that detects elements in an image that should attract attention. … GY’s worked on a model that uses contrasts to segment objects and BJ has a model that detects faces. We’ll use them as a base. … For now, most saliency models only detect objects and don’t pay attention to faces. There’s no ground truth for that. But what we say is that faces are also important and usually attract directly the attention. … And that’s the point: we want to include faces to saliency, basically.”

戈瑞：  “并且分割人脸。因为人脸检测器只输出矩形。......该模型可以有很多应用，例如显示或压缩。”

GY:  “And segment faces. Because face detectors output only rectangles. … There can be many applications [for the model], like in display or compression for example.”

该小组认为，显著性检测模型也应该考虑人脸，因为人脸在人类注意力机制中很重要。此外，将这一空白投入到显著性检测中将是一个很好的机会，可以合并该小组最近在低级分割和高级人脸检测方面的一些研究。将高级人脸检测与低级显著性检测相结​​合的想法源自以前的图像处理论文（Borji 2012；Karthikeyan、Jagadeesh 和 Manjunath 2013），灵感来自凝视预测（Cerf、Frady 和 Koch 2009）、认知心理学（Little、Jones 和 DeBruine 2011）和神经生物学（Dekowska、Kuniecki 和 Ja ś kowski 2008）的研究。但团队的目标是在刘等人 (2007) 之后，进一步推进王和李 (2008) 提出的显著性方向，提出一种能够检测和分割人脸轮廓的算法。为了实现如此精细的结果，GY 在分割方面所做的工作和 BJ 在人脸检测方面所做的工作将构成宝贵的资源。

According to the Group, saliency-detection models should also take human faces into account as faces are important in human attention mechanisms. Moreover, investing this interstice within saliency detection would be a good opportunity to merge some of the Group’s recent researches on both low-level segmentation and high-level face detection. The idea to combine high-level face detection with low-level saliency detection derived from previous image-processing papers (Borji 2012; Karthikeyan, Jagadeesh, and Manjunath 2013) inspired themselves by studies in gaze prediction (Cerf, Frady, and Koch 2009), cognitive psychology (Little, Jones, and DeBruine 2011), and neurobiology (Dekowska, Kuniecki, and Jaśkowski 2008). But the Group’s ambition here was to go further in the saliency direction as framed by Wang and Li (2008), after Liu et al. (2007), by proposing an algorithm capable of detecting and segmenting the contours of faces. In order to accomplish such subtle results, the previous work done by GY on segmentation and BJ on face detection would constitute a precious resource to work on.

该小组还希望构建一个可以有效处理更大范围的自然图像的显著性检测模型：

The Group also wanted to construct a saliency-detection model that could effectively process a larger range of natural images:

小组会议，实验室餐厅，2013 年 11 月 7 日

Group meeting, the Lab’s cafeteria, November 7, 2013

戈瑞：  “但你知道 [对 FJ 来说]，我们希望该算法能够检测多个物体和面孔。因为在显著性检测中，模型只能在简单图像上检测一两个物体。它们无法在复杂图像中检测多个显著物体。……但问题是，这没有基本事实。只有一两个物体的基本事实，而不是那么多面孔。”

GY:  “But you know [to FJ], we hope the algorithm could detect multiple objects and faces. Because in saliency detection, models can only detect like one or two objects on simple images. They don’t detect multiple salient objects in complex images. … But the problem is that there’s no ground truth for that. There’s only ground truth with like one or two objects, and not that many faces.”

在许多情况下，自然图像不仅仅捕捉到与清晰背景区分开的一两个物体；据该小组称，数码相机用户拍摄的照片通常比刘等人（2007）之后用于训练和评估显著性检测算法的照片更杂乱。事实上，至少在 2013 年 11 月，显著性检测已成为一个研究领域，其中算法只对那些具有清晰和不杂乱特征的罕见自然图像才越来越有效。但该小组也知道，这个问题与当时可用的显著性检测基本事实密切相关，这些事实都与刘等人最初对显著性的限制性定义有关，该定义仅适用于简单图像。从这一点来看，由于该小组想要提出一个可以检测不同且更微妙的显著性的模型，因此它必须构建这种显著性的目标；因为它想要提出一个模型，能够在更复杂、更真实的图像中计算和检测多个显著特征（物体和面部），所以它必须构建一个新的基本事实来收集复杂图像及其相应的多个显著特征。

In many cases, natural images not only capture one or two objects distinguished from a clear background; pictures produced by users of digital cameras—according to the Group—are generally more cluttered than those used to train and evaluate saliency-detection algorithms in the wake of Liu et al. (2007). Indeed, at least in November 2013, saliency detection was becoming a research area where algorithms were more and more efficient only on those—rare—natural images with clear and untangled features. But the Group also knew that this issue was intimately related to the then available ground truths for saliency detection that were all bound to Liu et al’s restricted initial definition of saliency that only fit simple images. From this point, as the Group wanted to propose a model that could detect a different and more subtle saliency, it had to construct the targets of such saliency; as it wanted to propose a model that could calculate and detect multiple salient features (objects and faces) in more complex and realistic images, it had to construct a new ground truth that would gather complex images and their corresponding multiple salient features.

该小组希望重新定义显著性问题，这一想法并非凭空而来。刘等人在 2007 年进行显著性研究时，计算机科学家很难针对复杂图像组织大规模社会调查。但在 2013 年 11 月，众包服务的日益普及带来了新的可能性：

The Group’s desire to redefine the terms of the saliency problem did not come ex nihilo. When Liu et al. did their research on saliency in 2007, it was difficult for computer scientists to organize large social surveys on complex images. But in November 2013, the growing availability of crowdsourcing services enabled new potentialities:

小组会议，实验室餐厅，2013 年 11 月 7 日

Group meeting, the Lab’s cafeteria, November 7, 2013

戈瑞：  “但我们想利用众包来得出新的基本事实，并要求人们标记他们认为突出的特征。……然后我们可以将其用于我们的模型并比较结果，你明白吗？”

GY:  “But we want to use crowdsourcing to do a new ground truth and ask people to label features they think are salient. … And then we could use that for our model and compare the results, you see?”

广义上讲，众包——“人群”和“外包”的缩写，最初由记者 Howe (2006) 提出——是“一种参与式在线活动，其中个人、机构、非营利组织或公司通过灵活的公开招募，向一群知识、异质性和数量各异的个人提议自愿承担一项任务”(Estellés - Arolas and González - Ladrón - de-Guevara 2012，195)。2013 年 11 月，亚马逊 (通过Amazon Mechanical Turk)、ClickWorker 或 Employment Crossing (通过S​​hortTask) 等多家公司提供了这项服务，这些公司自己的应用程序编程接口 (API) ¹⁰向主要位于美国和印度的注册在线临时工推荐调查。一旦工作者提交了他们完成的任务（时间和复杂程度可能有很大差异），设计调查的组织（例如研究机构、公司、个人）就可以决定其有效性。如果任务被认为有效，工作者将从众包公司获得最初在公开招募中指示的金额。如果任务被认为无效，工作者将一无所获，而且大多数情况下没有上诉的可能性。由于众包的道德经济最近成为批判社会学研究的对象，因此有必要为此写一篇简短的补充材料。

In broad strokes, crowdsourcing—a contraction of “crowd” and “outsourcing” initially coined by journalist Howe (2006)—is “a type of participative online activity in which an individual, an institution, a non-profit organization, or a company proposes to a group of individuals of varying knowledge, heterogeneity, and number, via a flexible open call, the voluntary undertaking of a task” (Estellés-Arolas and González-Ladrón-de-Guevara 2012, 195). In November 2013, this service was offered by several companies such as Amazon (via Amazon Mechanical Turk), ClickWorker, or Employment Crossing (via ShortTask), whose own application programming interfaces (APIs)¹⁰ recommended surveys to registered online contingent workers mainly located in the United States and India. Once a worker submits their completed task—which can vary greatly in time and complexity—the organization that designed the survey (e.g., a research institution, a company, an individual) can decide on its validity. If the task is considered valid, the worker receives from the crowdsourcing company the amount of money initially indicated in the open call. If the task is considered not valid, the worker receives nothing and has, most of the time, no possibility of appeal. As the moral economy of crowdsourcing has recently been the object of critical sociological studies, it is necessary to devote a short sidebar to it.

临时工长期以来一直支持着工业化进程。例如，正如 Pennington 和 Westover (1989) 所记录的那样，19 世纪 50 年代在英国发展起来的纺织业严重依赖场外制造业务，通常被称为“工业家庭作业”。生活在农村的妇女和儿童作为按需工人，被要求完成当时机器无法完成的关键收尾工作。几乎与此同时，美国也出现了类似的现象，特别是在宾夕法尼亚州的匹兹堡地区：尽管它经常被视为注定要消失的前工业时代的回忆，但与农村家庭合作组织起来的以佣金为基础的“计件工作”是扩大大规模制造规模的必要杠杆 (Albrecht 1982)。如果工会后来确实做到了，经过痛苦的斗争，企业在一定程度上改善了员工的工作条件（例如 1938 年美国的《公平劳动标准法》、1936 年法国的《马蒂尼翁协定》），这些改进主要涉及在指定生产现场进行的全职工作，这些工作大多仅供白人男性成年使用。因此，20 世纪上半叶对工薪阶层做出的让步主要涉及那些从可见性和接近性中受益的人：临时工分散、不引人注目、不受重视且被认为不需要技能，继续被忽视。除此之外——以及超出本侧边栏¹¹范围的许多其他事情——后来又增加了一项或多或少明确的企业战略，即规避工会和工作法规（这些法规已经为特定行业保留），其基础是信息和通信技术的日益普及。这种“工作场所裂变”策略（Weil 2014）与西方经济的金融化非常吻合，¹²有助于进一步促进外包：人们不再依赖受益于法定逻辑的员工，而是更愿意依赖全球各地的临时员工网络，这变得越来越可取和有价值。而众包，作为一种分布式计算机支持的按需低价值工作，可以看作是临时工作对工业资本主义的支持和修改的延续。正如 Gray 和 Suri（2019，58）所指出的那样：“如今这些按需工作是可消耗的幽灵工作的最新迭代。一方面，它们是当下必要的，但它们很容易被贬低，因为它们所做的任务通常被视为平凡或死记硬背，而经常被雇用来做这些工作的人没有文化影响力。” ¹³

Contingent work has long supported industrial efforts. As, for example, documented by Pennington and Westover (1989), the textile industry as it developed in England in the 1850s relied heavily on off-site manufacturing operations, often referred to as “industrial homework.” Women and children living in the countryside, operating as proto-on-demand workers, were asked to make crucial finishing touches too fine for the machines of the time. Almost simultaneously, a similar phenomenon was taking place in the United States, particularly in the Pittsburg, Pennsylvania, area: even though it was often seen as a reminiscence of a preindustrial era that was doomed to disappear, “piecework” organized on a commission basis in partnership with rural households was a necessary lever for the scaling up of mass manufacturing (Albrecht 1982). And if trade unions did later manage, through painful struggles, to somewhat improve the working conditions of employees (e.g., US Fair Labor Standards Act in 1938, French Accords de Matignon in 1936), these improvements mostly concerned full-time work carried out on designated production sites that was mostly reserved for white male adults. The concessions made to salaried workers during the first half of the twentieth century thus mostly concerned those who benefited from visibility and proximity: contingent work, which was scattered, not very visible, little valued, and considered unskilled, continued to pass under the radar. To this—and to many other things that are beyond the scope of this sidebar¹¹—was later added a more or less explicit corporate strategy of circumventing unionization and work regulations (which were already reserved for specific trades) based notably on the growing availability of information and communication technologies. This strategy of “fissuration of the workplace” (Weil 2014), well in line with the financialization of Western economies,¹² helped to further promote outsourcing: instead of depending on employees benefiting from statutory logic, it has become preferable and valued to depend on remote worldwide networks of contingent staff. And crowdsourcing, as distributed computer-supported on-demand low-valued work, can be seen as the continuation of contingent work’s support to and modification of industrial capitalism. As Gray and Suri (2019, 58) noted: “Those on-demand jobs today are the latest iteration of expendable ghost work. They are, on the one hand, necessary in the moment, but they are too easily devalued because the tasks that they do are typically dismissed as mundane or rote and the people often employed to do them carry no cultural clout.”¹³

让我们回到实验室。2013 年 11 月，与大多数人一样，该小组没有意识到当代众包流程所支持的普遍外包和临时工贬值的动态。这种无知可以从该小组经常使用的“用户”一词中找到，该术语指的是从事这种新形式的不稳定无产阶级的匿名工人^。14对于该小组来说，当时众包的估计收益是巨大的：一旦所需的 Web 应用程序被编码并设置了指令，例如“请突出显示直接吸引您注意的功能”，该小组就可以向众包公司付费，该公司的 API 将负责将调查与该小组 Web 应用程序的数十名低薪“用户”联系起来。反过来，这些“用户”——从现在起我将称之为“工人”——将为该小组提供调查问卷。Group 的服务器带有可在 Matlab 等软件包上处理的标记坐标。15^对于我们的故事来说，众包——一种相当容易获得的付费服务——带来了不同：与 2007 年刘等人的工作相比，Group 可以更轻松地收集许多手动标记的显著特征，并且至少在 2013 年 11 月，将显著性概念扩展到多个特征成为可能。

Let us come back to the Lab. In November 2013, like most people, the Group was not aware of the dynamics underlying generalized outsourcing and devaluation of contingent labor as supported by contemporary crowdsourcing processes. An indication of this unawareness could be found in the term “users” the Group often employed to refer to the anonymous workers engaged in this new form of precariat.¹⁴ For the Group, at that moment, the estimated benefits of crowdsourcing were huge: once the desired web application was coded and set with an instruction, such as “please highlight the features that directly attract your attention,” the Group would be able to pay a crowdsourcing company whose API would take charge of linking the survey to dozens of low paid “users” of the Group’s web application. In turn, these “users”—that I will from now on call “workers”—would feed the Group’s server with labeling coordinates that could be processed on software packages such as Matlab.¹⁵ For our story, crowdsourcing—as a rather easily available paid service—created a difference: the gathering of many manually labeled salient features became more manageable for the Group than it had been for Liu et al. in 2007, and an extension of the notion of saliency to multiple features became—at least in November 2013—doable.

众包带来的另一个不同是，显著性问题可能会被重新定义为连续的：

Another difference effected by crowdsourcing was a potential redefinition of the saliency problem as being continuous:

小组会议，实验室餐厅，2013 年 11 月 7 日

Group meeting, the Lab’s cafeteria, November 7, 2013

缩略词：  “那么，基本上你想要很多标签吗？”

FJ:  “So, basically you want many labels?”

戈瑞：  “是的，因为你知道，在最先进的人脸检测或显著性模型中，检测事物的方式只有二元性，比如有脸/没有脸、显著/不显著。我们还试图建立一个模型来评估人脸和物体的重要性并对其进行细分。比如‘这张脸比另一张脸更重要，而另一张脸又比那个物体更重要’等等。……但无论如何，要做到这一点（基于众包任务结果的地面实况），我们首先需要一个包含许多不同内容图像的数据集。”

GY:  “Yes because you know, in the state-of-the-art face detection or saliency models only detect things in a binary way, like face/no face, salient/not salient. What we also try to do is a model that evaluates the importance of faces and objects and segments them. Like ‘this face is more important than this other face which is more important than that object’ and so on. … But anyways, to do that [a ground truth based on the results of a crowdsourcing task], we first need a dataset with many images with different contents.”

氯：  “是的，我们考虑过至少 1,000 张图像来进行训练和评估。但这些图像必须包含不同的物体和不同大小的脸部。”

CL:  “Yes, we thought about something like 1,000 image at least, to train and evaluate. But it has to be images with different objects and faces with different sizes.”

戈瑞：  “我们必须选择图像；进行调查的好图像。……我们将尝试在春季提出一篇论文，所以我想，最好在一月份完成众包。”

GY:  “And we have to select the images; good images to run the survey. … We’ll try to propose a paper in [the] spring so it would be good to have finished crowdsourcing in January, I guess.”

如果用于构建基本事实的图像仅包含一个或两个对象，并且仅由几个人标记，则无法计算标记特征之间的关系值。从这一点来看，将显着性定义为二元问题（如刘等人（2007）的方式）是完全合理的。然而，由于该小组有能力发起一项社会调查，要求对包含许多特征的复杂图像的数据集进行许多标记，因此在方法论上可以为不同的标记特征分配相对重要性值。这是一个算术值的问题：如果一个特征被手动标记为显着，该小组只能获得一个二元值（前景和背景）。但如果许多工人将几个特征标记为或多或少显着，该小组可以获得连续的结果子集。简而言之，对于该小组来说，众包再次创造了不同，因为它使得创建具有相对连续值的新类型的目标成为可能。目前很难预测该集团的算法是否能够有效地接近这些微妙的结果。尽管如此，该集团想要构建的基本事实将通过提供模型应该以最佳方式尝试检索的目标来开发这种算法。

If the images used to construct the ground truth contained only one or two objects and were labeled only by several individuals, no relational values among the labeled features could be calculated. From this point, defining saliency as a binary problem in the manner of Liu et al. (2007) would make complete sense. Yet as the Group could afford to launch a social survey that asked for many labels on a dataset with complex images containing many features, it would become methodologically possible to assign relative importance values to the different labeled features. This was a question of arithmetic values: if one feature were manually labeled as salient, the Group could only obtain a binary value (foreground and background). But if several features were labeled as more or less salient by many workers, the Group could obtain a continuous subset of results. In short, for the Group, crowdsourcing once again created a difference by making it possible to create new types of targets with relatively continuous values. It was difficult at this point to predict if the Group’s algorithm would effectively be able to approach these subtle results. Nevertheless, the ground truth the Group wanted to constitute would enable the development of such an algorithm by providing the targets that the model should try to retrieve in the best possible way.

尽管该小组已在显著性检测和其他相关领域的前期工作基础上重新定义了显著性问题，但它仍然缺乏能够用数字方式确定这一新问题术语的基本事实：所需算法应处理的输入以及应尝试检索的输出（“目标”）仍需构建。从这个意义上说，该小组才刚刚开始问题化过程，这可能导致一种新的计算模型：其对显著性问题的新定义仍需配备（Vinck 2011）有形元素（一组新的复杂图像、一项众包任务、连续值、分割的人脸）以形成一个参考数据库，而该数据库反过来又构成了新计算机化计算方法的物质基础。借用 Michel Callon (1986) 的话，我们可以说，对于小组成员来说，新的基本事实是一个必需的切入点，可以使他们成为显著性检测研究界不可或缺的一员。如果没有新的基本事实，显著性检测模型仍将对不切实际的图像进行操作；它们仍将与一次性对象相关；它们仍将忽略人脸的检测和分割；因此，它们仍将与现实世界的应用无关。借助新的基本事实，该小组归因于显著性检测的这些缺点可能会得到克服。类似地——这次借用 Joan Fujimura (1987) 的观点——我们可以说，目前，该小组的显著性问题只能在实验室层面解决。该小组确实有时间和资金来开展这个项目，并且对如何运行它有深刻的见解。但是在没有任何基本事实的情况下，该小组没有切实可行的手段来与图像处理研究界和有效定义计算工作模型所需的具体任务阐明这个“实验室层面”。只有通过构建一个收集“输入数据”和“输出目标”的数据库，该小组才能够提出并最终发布一种能够解决该小组重新构建的显著性问题的算法。

Even though the Group had managed to build on previous works in saliency detection and other related fields to reframe the problem of saliency, it still lacked the ground truth that could numerically establish the terms of this new problem: both the inputs the desired algorithm should work on and the outputs (the “targets”) it should try to retrieve still needed to be constructed. In that sense, the Group was only at the beginning of the problematization process that may lead to a new computational model: its new definition of the saliency problem still needed to be equipped (Vinck 2011) with tangible elements (a new set of complex images, a crowdsourcing task, continuous values, segmented faces) to form a referential database that would, in turn, constitute the material base of the new computerized method of calculation. Borrowing from Michel Callon (1986), we might say that, for the members of the Group, the new ground truth appeared as an obligatory passage point that could make them become—perhaps—indispensable for the research community in saliency detection. Without a new ground truth, saliency-detection models would still operate on unrealistic images; they would still be one-off object related; they would still ignore the detection and segmentation of faces; and they would still, therefore, be irrelevant for real-world applications. With the help of a new ground truth, these shortcomings that the Group attributed to saliency detection may be overcome. In a similar vein—this time borrowing from Joan Fujimura (1987)—we might say that, at this point, the Group’s saliency problem was doable only at the level of its laboratory. The Group had indeed been given time and money to conduct the project and had insights on how to run it. But without any ground truth, the Group had no tangible means to articulate this “laboratory level” with both the research communities in image processing and the specific tasks required to effectively define a working model of computation. It is only by constructing a database gathering “input-data” and “output-targets” that the Group would be able to propose and, eventually, publish an algorithm capable of solving the saliency problem as the Group reframed it.

构建新的事实

Constructing a New Ground Truth

现在，我们对计算机科学家在尝试塑造新算法时有时会遇到的一些陷阱有了更好的认识。当我们跟踪该小组的显著性检测项目开始时，我们意识到，构建一个能够确立新研究方向的图像处理算法与塑造新的基本事实相伴而生，而新的基本事实应该精确地支持和装备算法的构建。然而，目前我们只考虑了该小组需要设计新基本事实的原因。但它实际上是如何做到的呢？

We have now a better sense of some of the pitfalls that sometimes get in the way of computer scientists trying to shape a new algorithm. As we were following the Group in the beginning of its saliency-detection project, we realized that the constitution of an image-processing algorithm capable of establishing a new research direction goes along with the shaping of a new ground truth that should precisely support and equip the constitution of the algorithm. Yet for now, we only considered the reasons why the Group needed to design a new ground truth. But how did it actually make it?

除了编写众包网络应用程序的代码外，该小组还于 2013 年 11 月和 12 月专门挑选了符合算法三个预期性能的图像：(1) 检测和分割显著特征（包括面部）的轮廓；(2) 在复杂图像中检测和分割这些显著特征；(3) 评估检测和分割的显著特征的相对重要性。这些规范促使小组专门组织了几次会议，讨论所选图像的内容和分布：

In addition to working on the coding of the crowdsourcing web application, the Group also dedicated November and December 2013 to the selection of images that echo the algorithm’s three expected performances: (1) detecting and segmenting the contours of salient features, including faces; (2) detecting and segmenting these salient features in complex images; and (3) evaluating the relative importance of the detected and segmented salient features. These specifications led to several Group meetings specifically organized to discuss the content and distribution of the selected images:

小组会议，实验室餐厅，2013 年 11 月 21 日

Group meeting, the Lab’s cafeteria, November 21, 2013

北京：  “好吧，我们可能会避免使用这种篮球照片，因为这些球员可能很有名。他们很好，因为篮球与面孔形成鲜明对比，但至少我认识一些球员。如果我认识，我们会添加其他特征，例如‘我认识这张脸’，所以我会给它贴上标签。”

BJ:  “Well, we may avoid this kind of basketball photo because these players may be famous-like. They are good because the ball contrasts with faces, but at least I know some of the players. And if I know, we include other features like ‘I know this face,’ so I label it.”

氯：  “我认为，如果某个人很有名，那么他的脸部的重要性就会增加，而我们只是想避免在我们的方法中对此进行建模。”

CL:  “I think maybe if you have somebody that is famous, the importance of the face increases and then we just want to avoid modeling that in our method.”

…

…

氯：  “好的。分布看起来好些了吗？”

CL:  “OK. And the distributions are looking better?”

缩略词：  “是的，绝对如此。BJ 只是告诉我需要改进的地方。”

FJ:  “Yes definitely. BJ just showed me what to improve.”

氯：  “好的。那么我们还需要考虑哪些其他变量？”

CL:  “OK. So what other variables do we consider?”

戈瑞：  “比如额叶等。但平衡它们真的很痛苦。”

GY:  “Like frontal and so on. But equalizing them is real pain.”

氯：  “但我们可以覆盖其中的一些；也许不能均衡。所以应该有正面图像，只有正面图像，然后是侧面图像，以及两者之间的混合图像。”

CL:  “But we can cover some of them; maybe not equalize. So there should be like the front face with images of just the front of the face and then there is the side face, and a mixture in between.”

选择过程花费了很长时间，因为必须收集各种各样的图像内容（如运动、肖像、侧面）来涵盖比其他基本事实更自然的情况。此外，内容中不应有会影响注意力过程的著名特征（如建筑物、喜剧演员、运动员）。我们可以看到，小组对算法的预期能力导向了这个手动选择过程：与刘等人（2007）的方法类似，但小组包括了更复杂的“自然情况”，数据集的组装由算法的未来任务驱动。16^到2013 年 12 月，已经收集了 800 张高分辨率图像（大部分来自 Flickr）并存储在实验室的服务器中。由于小组认为将人脸纳入显著性检测是该项目最重要的贡献，因此选定的图像中有 632 张包含人脸。

The selection process took time because a wide variety of image contents (e.g., sport, portraits, side faces) had to be gathered to cover more natural situations than the other ground truths. Also, no famous features (e.g., buildings, comedians, athletes) that could influence attention processes should be part of the content. As we can see, the Group’s anticipated capabilities for the algorithm oriented this manual selection process: similarly to Liu et al. (2007) but in a manner that made the Group include more complex “natural situations,” the assembling of a dataset was driven by the algorithm’s future tasks.¹⁶ By December 2013, eight hundred high-resolution images were gathered—mostly from Flickr—and stored in the Lab’s server. Since the Group considered the inclusion of faces within saliency detection as the most significant contribution of the project, 632 of the selected images included human faces.

在以问题为导向进行图像选择的同时，必须明确所选图像的组织工作，以免因文件数量增加和众包任务期间要收集的大量标记结果而负担过重。这种组织程序非常接近数据管理，意味着要实现一个全新的数据库，以便轻松检索和预测信息。此外，众包调查的形成也需要协调和调整：要问什么问题？如何收集和处理答案以实现项目的目标？这些都是至关重要的问题，因为通过众包获得的“原始”标记答案只能是矩形，而不是精确的轮廓：

In parallel to this problem-oriented selection of images, organizational work on the selected images had to be defined in order not to be overloaded by the increasing number of files and by the huge amount of labeled results to be gathered throughout the crowdsourcing task. This kind of organizational procedure was very close to data management and implied the realization of a whole new database for which information could be easily retrieved and anticipated. Moreover, the shaping of the crowdsourcing survey also required coordination and adjustments: What question would be asked? How would answers be collected and processed in order to fulfill the ambitions of the project? Those were crucial issues as the “raw” labeled answers obtained via crowdsourcing could only be rectangles and not precise contours:

小组会议，实验室餐厅，2013 年 12 月 12 日

Group meeting, the Lab’s cafeteria, December 12, 2013

氯：  “但是对于数据库，我们是否要重命名图像以保持一致性？”

CL:  “But for the database, do we rename the images so that we have a consistency?”

北京：  “嗯。……我不这么认为，因为现在我们可以通过文件 ID 追溯到网站。使用 Matlab，您可以将 jpg 文件存储在一个文件夹中并自动检索所有文件”

BJ:  “Hum. … I don’t think so because now we can track the files back to the website with their ID. And with Matlab you can like store the jpg files in one folder and retrieve all of them automatically”

…

…

氯：  “GY，您觉得怎么样？我们可以要求人们选择图像的某个区域，或者直接在该区域上进行分割之类的操作吗？”

CL:  “What do you think, GY? Can we ask people to select a region of the image or to do something like segmenting directly on it?”

戈瑞：  “我认为你无法通过众包获得像素级精度的答案。我们需要在实验室中实现像素级精度，因为如果我们问他们，这将是一项非常马虎的工作。或者无论如何都太慢而且成本太高。”

GY:  “I don’t think you can get pixel-precision answers with crowdsourcing. We’ll need to do the pixel-precision [in the Lab] because if we ask them, it’s gonna be a very sloppy job. Or too slow and expensive anyway.”

氯：  “那么你想要什么？这是你的 Matlab 代码来分割特征，对吗？”

CL:  “So what do you want? There is your Matlab code to segment features, right?”

戈瑞：  “是的，但那是低级的东西，像素精度 [分割]。我想，在我们收集坐标之后，这将会留到以后。无论如何，我仍然需要完成脚本 [以收集坐标]。真的很痛苦。...但我的想法是，就像要求人们在显著的东西上画矩形，然后收集带有其 ID 的坐标，然后使用这些信息推断出每个图像上哪个特征比其他特征更显著。显著特征的位置是一个非常模糊的决定，但切割边缘并不那么依赖。...你知道树的尽头在哪里，这就是我们想要的。没有人会过来说‘不！树在这里结束！’我想在大多数情况下，人与人之间的差异并不大。”

GY:  “Yes, but that’s low-level stuff, pixel-precision [segmentation]. It’s gonna be for later, after we collect the coordinates, I guess. I still need to finish the scripts [to collect the coordinates] anyway. Real pain. … But what I thought was just like ask people to draw rectangles on the salient things, then collect the coordinates with their ID and then use this information to deduce which feature is more salient than the other on each image. Location of the salient feature is a really fuzzy decision, but cutting up the edges is not that dependent. … You know where the tree ends, and that’s what we want. Nobody will come and say ‘No! The tree ends here!’ There is not so many variances between people I guess in most of the cases.”

氯：  “好的，那我们就来为矩形编写代码吧。如果这对用户来说很容易，那我们就这么做吧。”

CL:  “OK, let’s code for rectangles then. If that’s easy for the users, let’s just do that.”

选定图像的 ID 使该小组能够相当轻松地将图像放入 Matlab 数据库中。但在图像中，众包人员标记的显著特征更难处理，因为 GY 用于获取图像内容精确边界的交互式工具是基于低级信息的。因此，分割人脸等低对比度特征的边界可能需要几分钟，而负担得起的众包则涉及小而快的任务。该小组不能冒险收集“草率”的任务或花费不合理的资金来做这件事。17^因此，必须在实验室内对标记的特征进行后期处理以获得精确的轮廓。

The IDs of the selected images allowed the Group to put the images in a Matlab database rather easily. But within the images, the salient features labeled by the crowdworkers were more difficult to handle since GY’s interactive tool to get the precise boundaries of image contents was based on low-level information. As a consequence, segmenting the boundaries of low-contrasted features such as faces could take several minutes, whereas affordable crowdsourcing was about small and quick tasks. The Group could not take the risk of either collecting “sloppy” tasks or spending an infeasible amount of money to do so.¹⁷ The labeled features would thus have to be post-processed within the Lab to obtain precise contours.

此外，该项目的另一个潜在失败点在于众包网络应用程序的开发。事实上，让人们围绕特征绘制矩形，将这些矩形转换为坐标，并将它们存储到文件中以进行统计处理，这需要不简单的编程技能。到 2014 年 1 月，当众包网络应用程序全面投入运行时，它包含用 html、PHP 和 JavaScript 编写的七个不同的脚本（约七百行代码），这些脚本根据工作人员的输入相互响应（见图2.7 ）。然而，如果实验室的计算机科学家能够熟练掌握数值计算和 Matlab、C 或 C ++等编程语言，那么网页设计和社交汇集并不是他们必须接受的培训能力。

Moreover, another potential point of failure of the project resided in the development of the crowdsourcing web application. Indeed, asking people to draw rectangles around features, translating these rectangles into coordinates, and storing them into files to process them statistically required nontrivial programming skills. By January 2014, when the crowdsourcing web application was made fully operational, it comprised seven different scripts (around seven hundred lines of code) written in html, PHP, and JavaScript that responded to each other depending on the workers’ inputs (see figure 2.7). Yet, if the Lab’s computer scientists were at ease with numerical computing and programming languages such as Matlab, C, or C++, web designing and social pooling were not competencies for which they were necessarily trained.

编码和调试完成后（这本身就是一个微妙的过程（见第 4 章），不同的脚本被存储在实验室服务器的某个部分，该部分地址于 2014 年 1 月提供给现已解散的 ShortTask 公司，该公司的 API 提供了评分最高的临时工。到 2014 年 2 月，30 个工人的任务（数万个矩形坐标）以 .txt 文件的形式存储在小组数据库中，由于之前的准备步骤，这些文件已准备好进行处理。此时，先前收集的数据集的每一幅图像都与工人绘制的许多不同矩形相关联。通过在 Matlab 上叠加所有不同矩形的坐标，小组为每幅图像创建了一个具有不同强度的“权重图”，这些强度表明了对显著区域的相对一致性（见图2.8）。然后，小组对每幅图像应用一个广泛使用的阈值，该阈值取自 Otsu (1979)——Matlab 内部库的一部分——以仅保留被工人认为显著的权重区域。在第三步中，小组——实际上是 BJ 和我——花了整整两周的时间，手动分割显著区域内显著元素的轮廓以获得“显著特征”。最后，小组将显著区域图的平均值分配给相应的显著特征，以获得能够定义和评估新类型显著性检测算法的最终目标。这个艰苦的过程发生在 2014 年 2 月至 3 月之间；几乎一个月的时间都用于处理工人提供的坐标，然后由 html-JavaScript-PHP 脚本和数据库收集。

Once coded and debugged—a delicate process in its own right (see chapter 4)—the different scripts were stored in one section of the Lab’s server whose address was made available in January 2014 to the now-defunct company ShortTask whose API offered the best-rated contingent workers. By February 2014, thirty workers’ tasks qua tens of thousands of rectangles’ coordinates were stored in the Group’s database as .txt files, ready to be processed thanks to the previous preparatory steps. At this point, each image of the previously collected dataset was linked with many different rectangles drawn by the workers. By superimposing all the coordinates of the different rectangles on Matlab, the Group created for each image a “weight map” with varying intensities that indicated the relative consensus on salient regions (see figure 2.8). The Group then applied to each image a widely used threshold taken from Otsu (1979)—part of Matlab’s internal library—to keep only weighty regions that had been considered salient by the workers. In a third step that took two entire weeks, the Group—in fact, BJ and me—manually segmented the contours of the salient elements within the salient regions to obtain “salient features.” Finally, the Group assigned the mean value of the salient regions’ map to the corresponding salient features to obtain the final targets capable of defining and evaluating new kinds of saliency-detection algorithms. This laborious process took place between February and March 2014; almost a month was dedicated to the processing of the coordinates produced by the workers and then collected by the html-JavaScript-PHP scripts and database.

到 2014 年 3 月，该小组成功创建了具有相对显著性值的目标。然后可以将所选图像及其对应的目标组织成一个数据库，最终构成基本事实。从这一点来看，可以认为该小组有效地重新定义了显著性问题的术语：所需算法应进行的转换最终以数字方式定义。由于输入（所选图像）的定义和输出（目标）的定义，该小组终于拥有了一个可以通过数值计算解决的问题。

By March 2014, the Group successfully managed to create targets with relative saliency values. The selected images and their corresponding targets could then be organized as a single database that finally constituted the ground truth. From this point, one could consider that the Group effectively managed to redefine the terms of the saliency problem: the transformations the desired algorithm should conduct were—finally—numerically defined. Thanks to the definition of inputs (the selected images) and the definition of outputs (the targets), the Group finally possessed a problem that numerical computing could take care of.

当然，仅仅通过新的基本事实来建立问题的术语是不够的：为了提出一个实际的算法，该小组还必须根据他们刚刚建立的问题设计和编码指令列表，这些指令列表可以有效地将输入数据转换为输出目标。为了设计和编码这些指令列表，该小组随机从真实数据中选取了 200 幅图像作为训练集。在正式分析了该训练集的输入与目标之间的关系后，该小组提取了几个数值特征，这些特征虽然不是完全表达了这些输入与目标之间的关系。18^从训练集中提取和验证数值特征和参数，并将它们依次翻译成 Matlab 编程语言，整个过程花了将近一个月的时间。但在这个过程结束时，该小组拥有了一系列 Matlab 指令，这些指令能够将训练集的输入值转换为相对接近目标的值。

Of course, establishing the terms of a problem by means of a new ground truth was not enough: to propose an actual algorithm, the Group also had to design and code lists of instructions that could effectively transform input-data into output-targets according to the problem they had just established. To design and code these lists of instructions, the Group randomly selected two hundred images out of the ground truth to form a training set. After formal analysis of the relationships between the inputs and the targets of this training set, the Group extracted several numerical features that expressed—though not completely—these input-target relationships.¹⁸ The whole process of extracting and verifying numerical features and parameters from the training set and translating them sequentially into Matlab programming language took almost a month. But at the end of this process, the Group possessed a list of Matlab instructions that was able to transform the input values of the training set into values relatively close to those of the targets.

截至 2014 年 3 月底，该小组利用其剩余的真实数据库来评估该算法，并将其与已有算法进行比较显著性检测算法在精确度和召回率方面的比较（见图2.9）。这次对抗的结果令人满意，该小组算法的特点和性能最终在一份草稿中进行了总结，并提交给了重要的欧洲图像处理会议。

By the end of March 2014, the Group used the remainder of its ground-truth database to evaluate the algorithm and compare it with already available saliency-detection algorithms in terms of precision and recall measures (see figure 2.9). The results of this confrontation being satisfactory, the features and performances of the Group’s algorithm were finally summarized in a draft paper and submitted to an important European Conference on image processing.

正如这些小组会议和文件所示，只有在新定义的显著性问题得到人类工作者的解决并在真实数据库中表达后，小组的算法才能投入使用。从这个意义上说，能够解决新定义的显著性问题的 Matlab 指令列表的最终确定遵循了小组所从事的问题化过程。显著性的理论重构、Flickr 上特定图像的选择、网络应用程序的编码、Matlab 数据库的创建、处理工人的坐标：所有这些实践都是设计基本事实所必需的，最终允许提取算法的相关数字特征及其评估。当然，构建基本事实所需的平凡工作不足以完成复杂的 Matlab 指令列表，最终有效地处理图像的像素：关键的认证数学主张也需要以机器可读的格式表达和表达。然而，通过提供训练集来提取算法的数字特征，并通过提供评估集来衡量算法的性能，基本事实极大地参与了算法的完成。

As these Group meetings and documents show, the Group’s algorithm could only be made operational once the newly defined problem of saliency had been solved by human workers and expressed in a ground-truth database. In that sense, the finalization of Matlab lists of instructions capable of solving the newly defined problem of saliency followed the problematization process in which the Group was engaged. The theoretical reframing of saliency, the selection of specific images on Flickr, the coding of a web application, the creation of a Matlab database, the processing of the workers’ coordinates: all these practices were required to design the ground truth that ended up allowing the extraction of the relevant numerical features of the algorithm as well as its evaluation. Of course, the mundane work required for the construction of the ground truth was not sufficient to complete the complex lists of Matlab instructions that ended up effectively processing the pixels of the images: critical certified mathematical claims also needed to be articulated and expressed into machine-readable format. Yet, by providing the training set to extract the numerical features of the algorithm and by providing the evaluation set to measure the algorithm’s performances, the ground truth greatly participated in the completion of the algorithm.

上述要素并非微不足道，在继续前进之前，需要进行一些更深入的思考。2013 年 11 月，该小组只有少数几个要素可供利用。它有愿望（例如，对以前的论文提出质疑）、技能（例如，数学和编程能力）、手段（例如，可以查阅学术期刊、功能强大的计算机）和希望（例如，在图像处理领域有所作为）。但仅凭这些要素还不足以有效地塑造其新的预期算法。2013 年 11 月，该小组还需要一个可以作为基本基质的经验基础；它需要建立一种物质连贯性，以建立其未来模型的验证。这是新地面真相（更应该称为地面真相）的全部好处，因为现在可以找到并产生一组现象（这里是显着差异）作为分析参考。 2014 年 3 月，当这个文本固定得以实现后，该团体所居住的世界就不再是原来的样子了：它被数据库中物化的一组关系所丰富和定位。最终从这个数据库中产生的算法组织、再现并在某种意义上神圣化了嵌入其中的关系。从一个静态和特定的事实基础中产生了一个操作算法，该算法有可能以不同的配置再现和促进事实基础的组织规则。通过植根尚未构建的算法，该团体收集的事实基础将其算法的设计导向了特定的方向。从这个意义上说，新的事实基础是该团体算法的偶然但必要的偏见^。19

The above elements are not so trivial, and some deeper reflections are required before moving forward. In November 2013, the Group had only few elements at its disposal. It had desires (e.g., contesting previous papers), skills (e.g., mathematical and programming abilities), means (e.g., access to academic journals, powerful computers), and hopes (e.g., make a difference in the field of image processing). But these elements alone were not enough to effectively shape its new intended algorithm. In November 2013, the Group also needed an empirical basis that could serve as a fundamental substratum; it needed to ground a material coherence that could establish the veridiction of their future model. This was the whole benefit of the new ground truth—which should rather be called grounded truth—as it was now possible to found and bring into existence a set of phenomena (here, saliency differentials) operating as an analytical referential. Once this scriptural fixation was achieved in March 2014, the world the Group inhabited was no longer the same: it was enriched and oriented by a set of relations materialized in a database. And the algorithm that finally came out from this database organized, reproduced, and in a sense, consecrated the relations embedded in it. From a static and particular ground truth emerged an operative algorithm potentially capable of reproducing and promoting the organizational rules of the ground truth in different configurations. By rooting the yet-to-be-constructed algorithm, the ground truth as assembled by the Group oriented the design of its algorithm in a particular direction. In that sense, the new ground truth was the contingent yet necessary bias of the group’s algorithm.¹⁹

计算模型容易受到手动收集和处理的数据的约束，并因此而产生偏差，这种倾向并不局限于数字图像处理领域。例如，正如 Edwards (2013) 以气候学为例所表明的那样，繁琐的天气数据收集、标准化和汇编工作，以产生准确的地球气候基本事实，对于通用环流模型 (GCM) 的参数化和评估都至关重要。20^当然，就像在图像处理领域一样，气候学家构建的基本事实并不能保证定义准确有效的 GCM：还需要流体动力学、统计学和（并行）计算机编程方面的关键见解。然而，如果没有提供参数和评估的基本事实，就不会出现高效和值得信赖的 GCM。对于用于手写识别或垃圾邮件过滤的机器学习算法，Burrell (2016, 5–6) 指出了“测试数据”在设置这些算法的学习参数以及评估其性能方面的重要性。在这里，基本事实也显得至关重要，它定义了算法在统计上学习的内容，并允许评估其学习性能。²¹许多高频交易算法似乎也是如此：正如 MacKenzie (2014, 17–31) 所指出的，对以前金融交易的详细分析以及金融经济学的权威文献可以作为“执行”和“自营交易”算法的形成和评估的实证基础。

This propensity of computational models to be bound to and fundamentally biased by manually gathered and processed data is not limited to the field of digital image processing. For example, as Edwards (2013) showed for the case of climatology, the tedious collection, standardization, and compilation of weather data to produce accurate ground truths of the Earth’s climate is crucial for both the parametrization and evaluation of General Circulation Models (GCMs).²⁰ Of course, just as in the field of image processing, the construction of ground truths by climatologists does not guarantee the definition of accurate and effective GCMs: crucial insights in fluid dynamics, statistics, and (parallel) computer programming are also required. Yet, without ground truths providing parameters and evaluations, no efficient and trustworthy GCM could come into existence. For the case of machine learning algorithms for handwriting recognition or spam filtering, Burrell (2016, 5–6) noted the importance of “test data” in setting the learning parameters of these algorithms as well as in evaluating their performances. Here as well, ground truths appear central, defining what is statistically learned by algorithms and allowing the evaluation of their learning performances.²¹ The same seems also to be true of many algorithms for high-frequency trading: as MacKenzie (2014, 17–31) suggested, detailed analysis of former financial transactions as well as the authoritative literature of financial economics work as empirical bases for the shaping and evaluation of “execution” and “proprietary trading” algorithms.

然而，尽管经验证据越来越多，算法倾向于与显然不能简化为单纯数据集的真实数据库存在联系，但在有关算法的大量计算机科学文献中，这一点仍然很少被讨论。这个问题通常被忽略：数学分析和编程技术（有时非常复杂）是在构建、接受、分发和访问真实数据库之后才讨论的，就好像已经构建、接受、分发和访问了真实数据库一样。我在第 1 章中所说的算法标准概念的理论探索倾向于理所当然地认为存在稳定和共享的引用存储库。这种遗漏甚至可能正是使这种算法愿景成为可能的原因：将算法视为确保从问题到解决方案的计算机化转变的工具可能意味着假设已经定义的问题和已经可评估的解决方案。

Yet, despite growing empirical evidences, algorithms’ tendency to be existentially linked to ground-truth databases that cannot, obviously, be reduced to mere sets of data remains little discussed in the abundant computer science literature on algorithms. The issue is generally omitted: mathematical analysis and programming techniques, sometimes highly complex, are discussed after, or as if, a ground truth has been constructed, accepted, distributed, and made accessible. The theoretical exploration of what I called in chapter 1 the standard conception of algorithms tends to take for granted the existence of stable and shared referential repositories. This omission may even be what makes such a vision of algorithms possible: considering algorithms as tools ensuring the computerized transition from problems to solutions might imply to suppose already defined problems and already assessable solutions.

不过，一些社会学家（其中大多数是受 STS 启发的）确实在正面考虑这个话题。在对预测算法系统的批评中，Barocas 和 Selbst (2016) 警告了问题定义和训练集收集可能带来的有害后果。同样，Lehr和 Ohm (2017) 强调了统计学习算法设计中“玩弄数据”的手工方面。最近，Bechmann 和 Bowker (2019) 基于这些论点提出了价值问责设计的概念：呼吁系统性地努力使与算法相关的数据收集、准备和分类中涉及的任意选择更加明确。继 Ananny 和 Crawford (2018) 之后，他们因此建议，为了更好地理解算法行为，事前关注地面实况过程可能比事后审计或源代码审查更具结论性（例如，正如 Bostrom [2017] 和 Sandvig 等人 [2016] 所提出的）。类似地，Grosman 和 Reigeluth (2019) 研究了一种用于检测威胁行为的算法安全系统的设计。他们表明，算法需要解决的问题的定义（以及因此需要检测到的“真正值”）源自集体问题化过程，包括发起人之间的讨论和妥协、对法律文件的不同解释，以及项目工程师对威胁性和非攻击性行为的现场模拟。他们得出结论，算法系统特有的规范性也必须根据使这种规范性可表达的紧张关系来考虑。总之，所有上述作者都发现了与该小组刚刚经历的过程类似的过程。他们的调查还表明，所谓的“算法”通常源自集体过程，这些过程以偶然但必要的参考资料形式得到物质表达。

Some sociologists—most of them STS-inspired—do consider the topic head on, though. In their critique of predictive algorithmic systems, Barocas and Selbst (2016) warned against the potentially harmful consequences of problem definition and training sets’ collection. In a similar way, Lehr and Ohm (2017) emphasized on the handcrafted aspect of “playing with the data” for the design of statistical learning algorithms. More recently, Bechmann and Bowker (2019) built on these arguments to propose the notion of value-accountability-by-design: a call for systemic efforts to make arbitrary choices involved in algorithm-related data collection, preparation, and classification more explicit. In the wake of Ananny and Crawford (2018), they thus suggest that, to better appreciate algorithmic behavior, ex ante focus on ground-truthing processes might be more conclusive than ex post audits or source code scrutinization (as it is, for example, proposed in Bostrom [2017] and Sandvig et al. [2016]). In a similar way, Grosman and Reigeluth (2019) investigated the design of an algorithmic security system for the detection of threatening behaviors. They show that the definition of the problem that the algorithm will have to solve—and, therefore, the “true positives” it will have to detect—derive from collective problematization processes that include discussions and compromises among sponsors, competing interpretations of legal documents, and on-site simulations of threatening and inoffensive behaviors conducted by the project’s engineers. They conclude that the normativity proper to algorithmic systems must also be considered in the light of the tensions that contributed to making this normativity expressible. In sum, all the above-mentioned authors have uncovered processes that resemble the one the Group had just gone through. Their investigations also show that what is called an “algorithm” often derives from collective processes expressed materially in contingent, but necessary, referential repositories.

在当前研究的早期阶段，定义所有算法都具有的一般属性是不明智的。然而，基于本章的初步见解以及越来越多涉及类似问题的研究，我们可以合理地假设，我们喜欢称之为“算法”的许多实体背后都有真实数据库，这些数据库使设计者能够提取相关的数字特征并评估输入数据自动转换为输出目标的准确性。因此，一旦这些算法——一旦“进入野外”，在其生产场所之外——自动处理新数据，它们各自的初始真实数据——以及参与塑造它们的习惯、愿望和价值观——也会被调用，并在一定程度上得到促进。正如我将在本章末尾进一步阐述的那样，根据算法的执行效果研究这些算法与算法对我们生活日益增长的影响相比，这些算法试图检索的输出目标所构成的集体过程似乎是一个令人兴奋但仍未得到充分探索的研究课题。

At this early stage of the present inquiry, it would be unwise to define a general property common to all algorithms. Yet based on the preliminary insights of this chapter and the growing body of studies that touched on similar issues, one can make the reasonable hypothesis that behind many of these entities we like to call “algorithms” lie ground-truth databases that have made designers able to extract relevant numerical features and evaluate the accuracy of the automated transformations of inputs-data into output-targets. Consequently, as soon as such algorithms—once “in the wild,” outside of their production sites—automatically process new data, their respective initial ground truths—along with the habits, desires, and values that participated in their shaping—are also invoked and, to a certain extent, promoted. As I will further develop at the end of this chapter, studying the performative effects of such algorithms in the light of the collective processes that constituted the output-targets these algorithms try to retrieve appears a stimulating, yet still underexplored, research topic when compared with the growing influence algorithms have on our lives.

几乎被接受（但被拒绝）

Almost Accepted (Yet Rejected)

2014 年 6 月 19 日：审稿人拒绝了该小组的论文。该小组非常失望，几个月的精心工作没有得到一篇可以开启新研究方向并产生大量引用的论文的回报。但考虑到三位审稿人给出的理由，他们也感到不解和惊讶。

June 19, 2014: The reviewers rejected the Group’s paper. The Group was greatly disappointed to see several months of meticulous work unrewarded by a publication that could have launched new research lines and generated many citations. But the feeling was also one of incomprehension and surprise in view of the reasons provided by the three reviewers.

除了对将面部信息纳入显著性检测的实用性表示怀疑之外，审稿人还一致认为该小组的论文存在一个看似关键的缺陷：计算模型的性能比较仅针对该小组新的基本事实：

Along with doubts about the usefulness of incorporating face information within saliency detection, the reviewers agreed on one seemingly key deficiency of the Group’s paper: the performance comparisons of the computational model were only made with respect to the Group’s new ground truth:

指定审阅者 1

Assigned Reviewer 1

论文并未表明所提出的方法在公开基准测试中的表现也优于其他最先进的方法。……本文的实验评估仅针对自收集的人脸图像进行。更多的评估数据集将更有说服力。……需要进行更多实验来证明所提出的方法。

The paper does not show that the proposed method also performs better than other state-of-the-art methods on public benchmark ground truths. … The experiment evaluation in this paper is conducted only on the self-collected face images. More evaluation datasets will be more convincing. … More experiment needs to be done to demonstrate the proposed method.

指定审阅者 2

Assigned Reviewer 2

实验仅针对作者创建的基本事实进行测试。……如果对其他基本事实进行实验，并分别报告面部图像和非面部图像的结果，将会更有见地。这样，人们可以更彻底地评估面部重要性图的实用性。

The experiments are tested only on the ground truth created by the authors. … It would be more insightful if experiments on other ground truths were carried out, and results on face images and non-face images were reported, respectively. This way one can more thoroughly evaluate the usefulness of a face-importance map.

指定审阅者 3

讨论仍然过于主观，不足以支持其科学见解。从这个意义上讲，对现有数据集的评估很重要。

Assigned Reviewer 3

The discussion is still too subjective and not sufficient to support its scientific insights. Evaluation on existing datasets would be important in this sense.

审稿人认为该论文的技术方面是合理的。但他们质疑，新的最佳显著性检测模型（正如该小组在论文中提出的）是否只能与用于创建该模型的基本事实进行对比。事实上，为什么不将这个新模型与已经存在的显著性检测基本事实进行对比呢？如果该模型真的比已经发布的模型“更有效”，那么它也应该在用于塑造和评估先前发布的显著性检测模型性能的基本事实方面更有效。换句话说，由于小组将其模型与以前的模型进行比较，根据评审人员的说法，小组应该更彻底地比较其性能。但为什么小组在评估工作进行到一半时就停下来，只根据新的基本事实比较其模型呢？

The reviewers found the technical aspects of the paper to be sound. But they questioned whether the new best saliency-detection model—as the Group presented it in the paper—could be confronted only with the ground truth used to create it. Indeed, why not confront this new model with the already available ground truths for saliency detection? If the model were really “more efficient” than the already published ones, it should also be more efficient on the ground truths used to shape and evaluate the performances of the previously published saliency-detection models. In other words, since the Group presented its model as commensurable with former models, the Group should have—according to the reviewers—more thoroughly compared its performances. But why did the Group stop halfway through its evaluation efforts and compare its model only with respect to the new ground truth?

2014 年 6 月 19 日，在 CSF 餐厅露台上与 BJ 进行讨论

Discussion with BJ on the terrace of the CSF’s cafeteria, June 19, 2014

缩略词：  委员会不喜欢我们自己创造基本事实？²²

FJ:  The committee didn’t like that we created our own ground truth?²²

北京：  不，我的意思是，我们只是测试了这个，但没有测试其他的。

BJ:  No. I mean, it’s just that we tested on this one but we did not test on the other ones.

缩略词：  他们想让你根据已经存在的事实进行测试？

FJ:  They wanted you to test on already existing ground truths?

北京：  是的。

BJ:  Yes.

缩略词：  但你为什么不这么做呢？

FJ:  But why didn’t you do that?

北京：  好吧，这就是问题所在：为什么我们没有在其他数据集上测试它？我们有理由。我们的模型是关于人脸分割和多种特征的。但在其他数据集中，大多数人脸图像不超过十张。……在显著性区域，大多数人不研究人脸检测和多种特征。他们研究中间有汽车或鸟的图像。你总是会看到一只鸟或类似的东西。所以在这些数据集上测试我们的模型毫无意义。它们没有涵盖我们的模型所做的事情。……这就是问题所在：如果你做了经典的改进，你肯定会在大型会议上展示一些东西。但如果你提出新的东西，那么人们就会以某种方式误解这个概念。

BJ:  Well, that’s the problem: Why did we not test it on the others? We have a reason. Our model is about face segmentation and multiple features. But in the other datasets, most of them do not have more than ten face images. … In the saliency area, most people do not work on face detection and multiple features. They work on images where there is a car or a bird in the center. You always have a bird or something like this. So it just makes no sense to test our model on these datasets. They just don’t cover what our model does. … That’s the thing: if you do classical improvement, you are ensured that you will present something at big conferences. But if you propose new things, then somehow people just misunderstand the concept.

将该模型与之前的地面实况进行对比，对于该小组来说，在技术上并不困难；这些地面实况在网上可以免费获得，而且这种性能评估所需的 Matlab 脚本与用于生成图 2.9所示结果的脚本大致相同。该小组没有进行此类比较的主要原因是，从之前的地面实况中得出的先前模型肯定会获得更好的性能结果。由于该小组的模型并非为解决之前地面实况定义的显著性问题而设计的，因此它肯定会被这些地面实况的“原生”模型所超越。

It would not have been technically difficult for the Group to confront its model with the previous ground truths; they were freely available on the web, and such performance evaluations required roughly the same Matlab scripts as those used to produce the results shown in figure 2.9. The main reason the Group did not do such comparisons was that the previous models deriving from the previous ground truths would certainly have obtained better performance results. Since the Group’s model was not designed to solve the saliency problem as defined by the previous ground truths, it would certainly have been outperformed by these ground truths’ “native” models.

由于缺乏实证因素，我不会试图解释该小组为何觉得有必要围绕可量化绩效问题来构建其论文的论证思路。23^然而，根据本章的论点，我认为这次被拒绝的事件再次表明了图像处理算法如何与其基本事实紧密相关。算法从目标居中、对比对象的图像组成的地面实况中得出的算法将以某种方式设法检索这些目标。但是，当在目标为多个偏心对象和人脸的图像组成的地面实况上进行测试时，相同的算法很可能会产生统计上较差的结果。同样，从目标为多个偏心对象和人脸的图像组成的地面实况中得出的另一种算法将以某种方式设法检索这些目标。但是，当在目标为居中对比对象的图像组成的地面实况上进行测试时，它可能会产生统计上较差的结果。这两种算法都属于不同的类别；它们的局限性在于用于定义其作用范围的地面实况。正如 BJ 以一种戏剧性的方式所建议的那样，在一定程度上，我们得到了地面实况的算法。当算法从相同或非常相似的地面实况中得出时，它们可以在统计上比其他算法更有效。一旦两种算法从两个具有不同目标的地面实况中得出，它们就只能表现为不同。可以对不同的基本事实在方法论、数据选择、统计严谨性或工业潜力方面进行定性评估，但这两个计算模型本身存在不可约化的差异，并且无法比较。从本案例研究的角度来看（可能与审阅者的观点不同），该小组的致命错误可能是将绩效的定量改进与基本事实的定性改进混为一谈。

Due to a lack of empirical elements, I will not try to interpret the reasons why the Group felt obliged to frame the line of argument of its paper around issues of quantifiable performances.²³ Yet, in line with the argument of this chapter, I assume that this rejection episode shows again how image-processing algorithms can be bound to their ground truths. An algorithm deriving from a ground truth made of images whose targets are centered, contrastive objects will somehow manage to retrieve these targets. But when tested on a ground truth made of images whose targets are multiple decentered objects and faces, the same algorithm may well produce statistically poor results. Similarly, another algorithm deriving from a ground truth made of images whose targets are multiple decentered objects and faces will somehow manage to retrieve these targets. But when tested on a ground truth made of images whose targets are centered contrastive objects, it may well produce statistically poor results. Both such algorithms operate in different categories; their limits lie in the ground truths used to define their range of actions. As BJ suggested in a dramatic way, to a certain extent, we get the algorithms of our ground truths. Algorithms can be presented as statistically more efficient than others when they derive from the same—or very similar—ground truths. As soon as two algorithms derive from two ground truths with different targets, they can only be presented as different. Qualitative evaluations of the different ground truths in terms of methodology, data selection, statistical rigor, or industrial potentials can be conducted, but the two computational models themselves are irreducibly different and not commensurable. From the point of view of this case study—which may differ from the point of view of the reviewers—the Group’s fatal mistake might have been to mix up quantitative improvement of performances with qualitative refinement of ground truths.

有趣的是，在这次被拒事件发生一年后，该小组又提交了一篇论文，这次是提交给一个较小的图像处理会议。这篇论文的目的与之前被拒的论文完全相同：相同的基本事实和相同的计算模型。然而，该小组并没有强调其模型的统计性能，而是强调了其基本事实以及它允许​​在显著性检测中加入人脸分割的事实。在这篇获得会议“最佳短文奖”的第二篇论文中，计算模型被作为新基本事实应用潜力的一个例子来展示。

Interestingly, one year after this rejection episode, the Group submitted another paper, this time to a smaller conference in image processing. The objects of this paper were rigorously the same as those of the paper that was previously rejected: the same ground truth and the same computational model. Yet instead of highlighting the statistical performances of its model, the Group emphasized its ground truth and the fact that it allowed the inclusion of face segmentation within saliency detection. In this second paper that won the “Best Short Paper Award” of the conference, the computational model was presented as one example of the application potential of the new ground truth.

问题导向和/或公理化

Problem Oriented and/or Axiomatic

第一个案例研究只是一群年轻的计算机科学家在实验室。是否可以借鉴这一探索性案例研究的观察结果？我们能否利用其中的一些因素来提出更广泛的命题，并为本书以及其他未来可能对算法构成的探究勾勒出分析方向？我认为，这个案例研究不仅仅是关于一群年轻的计算机科学家和一个用于显着性检测的小型原型，它还充实了值得更深入探索的重要见解。那么，在本章的剩余部分，我将借鉴这个实证案例，尝试提出两个互补的算法社会学研究方向。

This first case study accounted for a small part of a four-month-long project in saliency detection run by a group of young computer scientists in the Lab. Is it possible to draw on the observations of this exploratory case study? Could we use some of the accounted elements to make broader propositions and sketch analytical directions for the present book as well as for other potential future inquiries into the constitution of algorithms? More than just concerning a group of young computer scientists and a small prototype for saliency detection, I think indeed that this case study fleshes out important insights that deserve to be explored more thoroughly. For the remaining part of this chapter then, I will draw on this empirical case to tentatively propose two complementary research directions for the sociological study of algorithms.

我认为这个案例研究隐含地提出了一种看待算法的新方法，这种方法在大幅扩展算法的标准定义的同时，仍然接受算法的标准定义。事实上，我们现在仍然可以将算法视为一组指令，用于计算解决给定的问题。尽管如第 1 章末尾所述，我有意不将算法的这个标准定义作为起点；在小组项目结束时，一旦从训练集中提取出数值特征并将其翻译成机器可读的语言，几个包含数千行指令的 Matlab 文件就构成了这样一个指令集。从这个角度来看，在理论层面研究这些指令集（例如，Knuth（1997a、1997b、1998、2011）、Sedgewick 和 Wayne（2011）、Dasgupta、Papadimitriou 和 Vazirani（2006）以及其他许多人提出的）与当前的问题完全相关。如何使用数学和机器可读语言以最有效的方式提出针对给定问题的解决方案确实是一个令人着迷的问题和研究领域。

I assume that this case study implicitly suggests a new way of seeing algorithms that still accepts their standard definition while expanding it dramatically. Indeed, we may now still consider an algorithm as being, at some point, a set of instructions designed to computationally solve a given problem. Though as explained at the end of chapter 1, I intentionally did not take this standard definition of algorithms as a starting point; at the end of the Group’s project, once the numerical features were extracted from the training set and translated into machine-readable language, several Matlab files with thousands of lines of instructions constituted just such a set. From that point of view, the study of these sets of instructions at a theoretical level—as proposed, for example, by Knuth (1997a, 1997b, 1998, 2011); Sedgewick and Wayne (2011); Dasgupta, Papadimitriou, and Vazirani (2006); and many others—is wholly relevant to the problem at hand. How to use mathematics and machine-readable languages in order to propose a solution to a given problem in the most efficient way is indeed a fascinating question and field of study.

但与此同时，我们也看到，算法要解决的问题并不是预先存在的：它必须在所谓的“问题化过程”中产生——一系列旨在经验性地定义要解决的问题的术语的集体实践。在我们的案例研究中，该小组首先借鉴了认知生物学权威期刊上发表的最新主张，将显着性问题重新定义为与面部相关且连续的。正如我们所看到的，该小组问题化过程的第一步意味着平凡而有问题的实践，例如对先前研究结果的批评（我们的对手错过了什么？）和纳入实验室的一些最新项目（如何开展我们的最新发展？）。该小组问题化过程的第二步意味着构建一个基础事实，以便将重新定义的显著性问题付诸实践。第二步还意味着一些平凡而又有问题的做法，例如在 Flickr 上收集数据集（我们选择什么图片？）、组织数据库（我们如何组织数据？）、设计众包任务（我们向工作人员提出什么问题？）以及处理结果（我们如何从矩形中获取特征轮廓？）。只有在这个过程的最后——一旦辛苦构建的目标与辛苦构建的数据集相关联以形成最终的基础事实数据库——该小组才能够制定、编程和评估一组 Matlab 指令，这些指令能够通过数值计算技术将输入转换为输出。简而言之，要设计一种能够解决新显著性问题的计算机化计算方法，该小组首先必须定义这个新问题的边界。

At the same time, however, we saw that the problem an algorithm is designed to solve does not preexist: it has to be produced during what one may call a “problematization process”—a succession of collective practices that aim to empirically define the terms of a problem to be solved. In our case study, the Group first drew on recent claims published in authoritative journals of cognitive biology to reframe the saliency problem as being face-related and continuous. As we saw, this first step of the Group’s problematization process implied mundane and problematic practices such as the critique of previous research results (what did our opponents miss?) and the inclusion of some of the Lab’s recent projects (how to pursue our recent developments?). The second step of the Group’s problematization process implied the constitution of a ground truth that could operationalize the reframed problem of saliency. This second step also implied mundane and problematic practices such as the collection of a dataset on Flickr (what images do we choose?), the organization of a database (how do we organize our data?), the design of a crowdsourcing task (what question do we ask to the workers?), and the processing of the results (how do we get contours of features from rectangles?). Only at the very end of this process—once the laboriously constructed targets have been associated to the laboriously constructed dataset in order to form the final ground-truth database—was the Group able to formulate, program, and evaluate the set of Matlab instructions capable of transforming inputs into outputs by means of numerical computing techniques. In short, to design a computerized method of calculation that could solve the new saliency problem, the Group first had to define the boundaries of this new problem.

从这些经验要素中，似乎出现了两种关于小组算法的互补观点。第一种观点可能认为小组算法是一组指令，旨在以最佳方式计算解决新问题。这种关于小组算法的传统观点反过来会强调小组用于将新基本事实的输入数据转换为其相应的输出目标的数学选择、制定实践和编程程序。小组如何操纵其训练集以提取此类任务的相关数值特征？小组如何将数学运算转化为代码行？它是否产生了最有效的结果？简而言之，这种对小组算法的看法将根据其计算特性对其进行分析。然而，对称地，对小组算法的第二种观点可能认为它是一组指令，旨在以最佳方式计算检索在特定问题化过程中设计的输出目标。反过来，对 Group 算法的第二种看法将重点放在导致定义该算法旨在解决的问题术语的具体情况和实践上。问题是如何定义的？数据集是如何收集的？众包任务是如何进行的？简而言之，本章所认可的第二种观点将分析 Group 算法与它最初得出的（并由此产生偏见的）基本事实的构建过程。

From these empirical elements, two complementary perspectives on the Group’s algorithm seem to emerge. A first perspective might consider the Group’s algorithm as a set of instructions designed to computationally solve a new problem in the best possible way. This first traditional view on the Group’s algorithm would, in turn, put the emphasis on the mathematical choices, formulating practices, and programming procedures the Group used to transform the input-data of the new ground truth into their corresponding output-targets. How did the Group manipulate its training set to extract relevant numerical features for such a task? How did the Group translate mathematical operations into lines of code? And did it lead to the most efficient result? In short, this take on the Group’s algorithm would analyze it in the light of its computational properties. Yet symmetrically, a second view on the Group’s algorithm might consider it as a set of instructions designed to computationally retrieve, in the best possible way, output-targets that were designed during a specific problematization process. This second take on the Group’s algorithm would, in turn, put the emphasis on the specific situations and practices that led to the definition of the terms of the problem the algorithm was designed to solve. How was the problem defined? How was the dataset collected? How was the crowdsourcing task conducted? In short, this second perspective—which this chapter endorsed—would analyze the Group’s algorithm vis-à-vis the construction process of the ground truth it originally derived from (and by which it was biased).

如果我们尝试性地扩展上述命题，我们最终会得到两种考虑算法的方法，它们都围绕着这些被称为基本事实的物质对象。我们可以称之为算法的公理视角，它将算法视为指令集，旨在以最佳方式计算解决由给定基本事实定义的问题。第二种互补的面向问题的算法视角将算法视为指令集，旨在计算检索在特定问题化过程中被定义为输出目标的内容。

If we tentatively expand the above propositions, we end up with two ways of considering algorithms that both pivot about these material objects called ground truths. What we may call an axiomatic perspective on algorithms would consider algorithms as sets of instructions designed to computationally solve in the best possible way a problem defined by a given ground truth. A second, and complementary, problem-oriented perspective on algorithms would consider algorithms as sets of instructions designed to computationally retrieve what has been defined as output-targets during specific problematization processes.

虽然我确实认为公理化和面向问题的算法观点是互补的，因此应该紧密结合——具体数值特征由基本事实所暗示（反之亦然）——但我也相信它们会导致不同的分析努力。通过考虑手头问题的条件，考虑算法的公理化方式有助于研究实际的数学和编程程序，这些程序最终会以最佳方式有效地将输入值集转换为输出值集。这听起来像是一个显而易见的陈述，但定义一种计算方法需要对方法的初始条件和预期结果达成最低限度的一致（Ritter 1995）。通过假设将输入数据转换为输出目标是可取的、相关的和可证明的，可以提出一个描述这种转换的分步模式。在计算机科学的情况下，可以探索、调整和利用具有许多不同认证规则和定理的不同数学领域，以最大限度地自动化从选定的输入数据到指定的输出目标的转变；图像处理中的线性代数（Klein 2013）、数据压缩中的概率论（Pu 2005）、数据结构中的图论（Tarjan 1983）、密码学中的数论（Koblitz 2012），或广受欢迎的机器学习程序中的统计学（和概率），据称适用于所有活动领域（Alpaydin 2016）。正如我们将在第 5 章和第 6 章中看到的那样，对这些不同的经过认证的数学知识体系的探索和教学必须尊重它们的本质：强大的运算符允许将真实输入数据可靠地转换为其相应的输出目标。

While I do think that both axiomatic and problem-oriented perspectives on algorithms are complementary and should thus be intimately articulated—specific numerical features being suggested by ground truths (and vice versa)—I also believe that they lead to different analytical efforts. By considering the terms of the problem at hand as given, the axiomatic way of considering algorithms facilitates the study of the actual mathematical and programming procedures that effectively end up transforming input sets of values into output sets of values in the best possible ways. This may sound like an obvious statement, but defining a calculating method requires minimal agreement on the initial terms and prospected results of the method (Ritter 1995). It is by assuming that the transformation of the input-data into the output-targets is desirable, relevant, and attestable that a step-by-step schema describing this transformation might be proposed. In the case of computer science, different areas of mathematics with many different certified rules and theorems can be explored, adapted, and enrolled to automate at best the passage from selected input-data to specified output-targets; linear algebra in the case of image processing (Klein 2013), probability theory in the case of data compression (Pu 2005), graph theory in the case of data structure (Tarjan 1983), number theory in the case of cryptography (Koblitz 2012), or statistics (and probabilities) in the case of the ever-popular machine-learning procedures supposedly adaptable to all fields of activity (Alpaydin 2016). As we will see in chapters 5 and 6, the exploration and teaching of these different certified mathematical bodies of knowledge must therefore be respected for what they are: powerful operators allowing the reliable transformative computation of ground-truth’s input-data into their corresponding output-targets.

如果以问题为导向的算法视角可能不直接关注算法的形成和计算有效性，那么它可能有助于更好地记录配置这些算法试图解决的问题术语的过程。将算法视为检索实体可能会强调参考数据库，这些数据库定义了算法试图检索和重现的内容；它们为了表达其真实性而建立的偏见。什么基本事实定义了该算法试图解决的问题的术语？这个基本事实数据库是如何构成的？什么时候？由谁构成的？通过指出在基本事实数据库中构成要检索的输出的时刻和位置，这种对算法的分析——Bechmann 和 Bowker (2019) 以及 Grosman 和 Reigeluth (2019) 为之做出的贡献——可能提出与算法及其设计者互动的新方式。这条研究途径仍处于起步阶段，此外，它可以将其结果与我在引言中提到的更明确的关键立场联系起来。如果 Noble (2018) 对搜索引擎谷歌所助长的种族主义刻板印象的研究，或 O'Neil (2016) 对专有评分算法所使用的代理如何倾向于惩罚最贫困人群的研究有效地起到了警示作用，那么改变当前状况的实际方法仍然需要进一步阐述。这时，作为本研究基石的“组合”概念再次发挥作用：在（合理的）愤慨之时，必须进行建设性的对抗，这本身意味着能够真实地展示自己。只要算法构成所涉及的实际工作仍然抽象和不确定，改变这项工作的生态就会仍然极其困难。从这个意义上说，改变根植于算法的偏见以使其促进不同的价值观可以通过使算法真实性背后的工作实践更加明显来实现。如果更多的研究能够探究算法所源自的真实实践，那么实际的组合潜力可能会慢慢被揭示出来。

If the problem-oriented perspective on algorithms may not directly focus on the formation and computational effectiveness of algorithms, it may contribute to better documenting the processes that configure the terms of the problems these algorithms try to solve. Considering algorithms as retrieving entities may put the emphasis on the referential databases that define what algorithms try to retrieve and reproduce; the biases they build on in order to express their veracity. What ground truth defined the terms of the problem this algorithm tries to solve? How was this ground-truth database constituted? And when? And by whom? By pointing at moments and locations where outputs to be retrieved were, or are, being constituted within ground-truth databases, this analytical look at algorithms—that Bechmann and Bowker (2019) and Grosman and Reigeluth (2019) contributed to igniting—may suggest new ways of interacting with algorithms and those who design them. This avenue of research, which is still in its infancy, could moreover link its results to those of the more explicitly critical positions I mentioned in the introduction. If the investigations by Noble (2018) on the racist stereotypes promoted by the search engine Google or by O’Neil (2016) on how proxies used by proprietary scoring algorithms tend to punish the poorest have effectively acted as warning signs, practical ways to change the current situation still need to be elaborated. This is where the notion of composition, the keystone of this inquiry, comes again into play: at the time of (legitimate) indignation, the time of constructive confrontation must follow, which itself implies being able to present oneself realistically. As long as the practical work subtending the constitution of algorithms remains abstract and indefinite, modifying the ecology of this work will remain extremely difficult. Changing the biases that root algorithms in order to make them promote different values may, in that sense, be achieved by making the work practices that underlie algorithms’ veracities more visible. If more studies could inquire into the ground-truthing practices algorithms derive from, then actual composition potentials may slowly be suggested.

第一部分即将结束。让我快速回顾一下到目前为止介绍的内容。在第 1 章中，我介绍了本研究的主要背景：一个学术实验室，我决定称之为“实验室”，其成员花费大量时间和精力组装和发布新的图像处理算法，从而以自己的水平参与计算机科学行业的异构网络。我还考虑了方法论问题并批判性地讨论了专业文献中普遍提出的算法概念。

Part I is now coming to an end. Let me then quickly recap the elements presented so far. In chapter 1, I presented the main setting of this inquiry: an academic laboratory I decided to call the “Lab” whose members spend a fair amount of time and energy assembling and publishing new image-processing algorithms, thus participating—at their own level—in the heterogeneous network of computer science industry. I also considered methodological issues and critically discussed the notion of algorithm as it is generally presented in the specialized literature.

在第 2 章中，我们深入研究了实验室的日常工作，并跟踪了一群年轻的计算机科学家，他们试图为图像处理领域的一个重要会议设计一种新算法。我们最初在实验室餐厅与该小组的接触起初令人困惑，但在快速浏览了有关显著性检测的图像处理文献后，我们能够理解为什么该小组的项目意味着建立一个新的参考数据库，该数据库可以定义其所需算法稍后应尝试解决的问题的术语。当我们解释这些平凡但至关重要的地面实况实践时，我们意识到一些对计算机科学从业者来说非常平庸但对其他许多人来说却令人惊讶的事情：事实证明，在一定程度上，我们得到了地面实况的算法。由于图像处理算法的构建意味着形成训练集以制定输入图像和输出目标之间的关系，以及形成评估集以测量和比较这些制定的关系的性能，因此图像处理算法（以及可能许多其他算法）必须以某种方式依赖于手动构建的精确提供这两个集合的地面真相。这一半发现进一步提出了一个研究议程，即两种互补的算法分析视角可以灌溉。首先，在第二章之后，“面向问题的视角”可以探索导致地面真相形成和传播的集体过程。这种对算法的非常规审视可能有助于为与数据正义和算法公平性相关的更广泛主题提供依据。然而，为了避免将算法简化为它们所衍生的地面真相，这种算法研究应该与算法的“公理视角”紧密结合，后者可以进一步探索从已经构成的基础事实中制定和评估计算模型。

In chapter 2, we dived into the daily work of the Lab and followed a group of young computer scientists trying to design a new algorithm for an important conference in image processing. Our initial encounter with the Group at the Lab’s cafeteria was at first confusing, but after a quick detour via the image-processing literature on saliency detection, we were able to understand why the Group’s project implied the shaping of a new referential database that could define the terms of the problem its desired algorithm should later try to solve. As we were accounting for these mundane yet crucial ground-truthing practices, we realized something very banal for practitioners of computer science but surprising to many others: it turns out that, to a certain extent, we get the algorithms of our ground truths. As the construction of image-processing algorithms implies the formation of training sets for formulating the relationships between input-images and output-targets as well as the formation of evaluation sets for measuring and comparing the performances of these formulated relationships, image-processing algorithms—and potentially many others—must rely, in one way or another, on manually constructed ground truths that precisely provide both sets. This half-discovery further suggested a research agenda that two complementary analytical perspectives on algorithms could irrigate. First, and in the wake of this chapter 2, a “problem-oriented perspective” could explore the collective processes leading to the formation and circulation of ground truths. This unconventional glance on algorithms may contribute to equipping broader topics related to data justice and algorithmic fairness. Yet to avoid reducing algorithms to the ground truths from which they derive, such studies of algorithms should be intimately articulated with an “axiomatic perspective” on algorithms that could further explore the formulation and evaluation of computational models from already constituted ground truths.

笔记

Notes

1. 1.本章扩展了 Jaton (2017)。感谢 Geoffrey Bowker、Roderic Crooks 和 John Seberger 就其中的一些主题进行的富有成效的讨论。

2. 1.  This chapter expands Jaton (2017). I thank Geoffrey Bowker, Roderic Crooks, and John Seberger for fruitful discussions about some of its topics.

3. 2.引号内的摘录是从录音中逐字转录的，略作修改以方便阅读。未引号内的摘录是从即兴记录的书面笔记中重新转录的。

4. 2.  Excerpts in quotes are literal transcriptions from audio recordings, slightly reworked for reading comfort. Excerpts not in quotes are retranscriptions from written notes taken on the fly.

5. 3.在第三章中，我批判性地讨论了许多认知研究所依赖的思维计算隐喻。

6. 3.  In chapter 3, I critically discuss the computational metaphor of the mind on which many cognitive studies rely.

7. 4.对注意力的研究早在 20 世纪 70 年代之前就已开始，尤其是通过 Neisser (1967) 的开创性研究，他提出了人类视觉处理系统中存在前注意阶段。

8. 4.  Studies on attention had already been engaged before the 1970s, notably through the seminal work of Neisser (1967) who suggested the existence of a pre-attentive stage in the human visual processing system.

9. 5 . 另一个重要的选择性注意方法的神经生物学模型是由 Wolfe、Cave 和 Franzel (1989) 提出的。这种选择性注意方法模型后来启发了竞争性的低级特征计算模型 (例如，Tsotsos 1989；Tsotsos 等 1995)。

10. 5.  Another important neurobiological model of selective attention method was proposed by Wolfe, Cave, and Franzel (1989). This model of selective attention method later inspired competing low-level feature computational models (e.g., Tsotsos 1989; Tsotsos et al. 1995).

11. 6.基于低级特征进行计算的算法类别很快成为自动驾驶汽车开发中关注的焦点，因为自动驾驶汽车需要实时图像寻求处理（Baluja 和 Pomerleau 1997；Grimson 1986；Mackworth 和 Freuder 1985）。

12. 6.  The class of algorithms that calculates on low-level features quickly became interesting for the development of autonomous vehicles for which real-time image processing was sought (Baluja and Pomerleau 1997; Grimson 1986; Mackworth and Freuder 1985).

13. 7.不同的高级检测算法仍然可以作为模块组装在同一个程序中，例如，可以检测人脸、汽车和狗等等。

14. 7.  Different high-level detection algorithms can nonetheless be assembled as modules in one same program that could, for example, detect faces and cars and dogs, and so on.

15. 8。当时，只有两种显著性检测算法被发表，分别是 Itti、Koch 和 Niebur (1998) 和 Ma 和 Zhang (2003)。但用于设计和评估这些算法的地面实况与实验室认知科学中使用的地面实况类似。这些地面实况的图像例如是被垂直划线打断的点集。因此，如果这前两种显著性检测算法当然可以处理自然图像，那么就无法对它们在此类图像上的性能进行评估。

16. 8.  At that time, only two saliency-detection algorithms were published, in Itti, Koch, and Niebur (1998) and Ma and Zhang (2003). But the ground truths used for the design and evaluation of these algorithms were similar to those used in laboratory cognitive science. The images of these ground truths were, for example, sets of dots disrupted by a vertical dash. As a consequence, if these first two saliency-detection algorithms could, of course, process natural images, no evaluations of their performances on such images could be conducted.

17. 9.计算机科学实验室收集的真相通常以可重复研究的名义在线提供（Vandewalle、Kovacevic 和 Vetterli 2009）。这种免费访问的对应内容是对这些真相首次提出的论文的正确引用。

18. 9.  Ground truths assembled by computer science laboratories are generally made available online in the name of reproducible research (Vandewalle, Kovacevic, and Vetterli 2009). The counterpart to this free access is the proper citation of the papers in which these ground truths were first presented.

19. 10.从最广泛的意义上讲，API 是一组通信协议，充当多个计算机程序之间的接口。如果 API 可以采用多种不同的形式（例如，硬件设备、Web 应用程序、操作系统），则其主要功能是稳定和黑盒化元素，以便可以在其之上构建其他元素。

20. 10.  An API, in its broadest sense, is a set of communication protocols that act as an interface among several computer programs. If APIs can take many different forms (e.g., hardware devices, web applications, operating systems), their main function is to stabilize and blackbox elements so that other elements can be built on top of them.

21. 11.有关临时工作的简要历史，请参阅 Gray 和 Suri (2019, 48–63)。有关众包对当代资本主义的影响，另请参阅 Casilli (2019)。

22. 11.  For a condensed history of contingent work, see Gray and Suri (2019, 48–63). On what crowdsourcing does to contemporary capitalism, see also Casilli (2019).

23. 12.正如 Gray 和 Suri (2019, 55–56) 所说：“遵循一种基本上未经检验的管理理论，20 世纪 80 年代的一波企业削减了所有可以定义为‘非必要业务运营’的业务——从清洁办公室到调试软件程序——以便向股东展示其真正的价值，即‘投资回报率’（行业术语为 ROI）和‘核心竞争力’。 ……股东们奖励那些愿意通过外包来削减成本和减少全职员工数量的企业。”

24. 12.  As Gray and Suri (2019, 55–56) put it: “Following a largely untested management theory, a wave of corporations in the 1980s cut anything that could be defined as ‘non-essential business operations’—from cleaning offices to debugging software programs—in order to impress stockholders with their true value, defined in terms of ‘return on investment’ (in industry lingo, ROI) and ‘core competencies.’ … Stockholders rewarded those corporations that were willing to use outsourcing to slash costs and reduce full-time-employee ranks.”

25. 13.但值得注意的是，按需工作并不一定会让人疏远。正如 Gray 和 Suri (2019, 117) 指出的那样：“当工人的需求和市场需求得到适当的调整和匹配时，[按需工作] 可以转变为更实质性和更令人满意的工作。如果不加以控制或隐藏在软件背后，而不是被视为一个快速增长的全球就业世界，它可能会迅速转变为幽灵工作。” 全国家庭工人联盟及其“良好工作守则”质量标签提出了使众包更具可持续性的具体方法。关于这个主题，请参阅 Scheiber (2016)。

26. 13.  It is important to note, however, that on-demand work is not necessarily alienating. As Gray and Suri (2019, 117) noted: “[on-demand work] can be transformed into something more substantive and fulfilling, when the right mixture of workers’ needs and market demands are properly aligned and matched. It can rapidly transmogrify into ghost work when left unchecked or hidden behind software rather than recognized as a rapidly growing world of global employment.” Concrete ways to make crowdsourcing more sustainable have been proposed by the National Domestic Workers Alliance and their “Good Work Code” quality label. On this topic, see Scheiber (2016).

27. 14.然而，这种对众包底层流程的共同无知可能因身份原因而受到重视和维持，正如 Irani (2015, 58) 所指出的：“将工人转变为计算服务……不仅满足了雇主的劳动力需求和经济利益，也满足了他们保持首选身份的愿望；也就是说，雇主不再将自己理解为信息工厂的管理者，而是继续将自己视为备受推崇的程序员、企业家和创新者。”

28. 14.  However, this shared unawareness toward the underlying processes of crowdsourcing may be valued and maintained for identity reasons, for as Irani (2015, 58) noted: “The transformation of workers into a computational service … serves not only employers’ labor needs and financial interests but also their desire to maintain preferred identities; that is, rather than understanding themselves as managers of information factories, employers can continue to see themselves as much-celebrated programmers, entrepreneurs, and innovators.”

29. 15. Matlab 是一款私人持有的数值计算数学软件，其核心是其自有的解释型高级编程语言。由于 Matlab 能够灵活地解决线性代数问题（将所有整数视为标量），因此被广泛应用于计算机科学、电子工程和经济学领域的研究和工业用途。然而，由于 Matlab 主要使用解释型编程语言（例如 Python 语言，它是 Matlab 在应用研究方面的主要竞争对手），因此必须通过解释器将其程序翻译成机器可读的二进制代码，才能使硬件有效地计算数据。与直接用 C 或 C ++等编译型语言编写的程序相比，这种复杂的解释步骤使其处理重矩阵的效率较低。有关 Matlab 的简要历史，请参阅 Haigh (2008)。

30. 15.  Matlab is a privately held mathematical software for numerical computing built around its own interpreted high-level programming language. Because of its agility to design problems of linear algebra—all integers being considered scalars—Matlab is widely used for research and industrial purposes in computer science, electrical engineering, and economics. Yet, as Matlab works mainly with an interpreted programming language—just like the language Python that is now Matlab’s main competitor for applied research purposes—its programs have to be translated into machine-readable binary code by an interpreter in order to make the hardware effectively compute data. This complex interpretative step makes it less efficient for processing heavy matrices than, for example, programs directly written in compiled languages such as C or C++. For a brief history of Matlab, see Haigh (2008).

31. 16.在第6章中，我们将更深入地考虑地面实况和制定活动之间的关系。

32. 16.  In chapter 6, we will more thoroughly consider the relationship between ground-truthing and formulating activities.

33. 17 . 众包公司的服务花费了实验室大约 950 美元。

34. 17.  The services of the crowdsourcing company costed the Lab around US$950.

35. 18 . 从训练集提取的数值特征与“二维高斯函数”、“空间紧凑性”、“基于对比度的过滤”、“高维高斯滤波器”和“元素唯一性”等相关。在第 6 章中，我将以“二维高斯函数”为例，讨论这些公式化实践。

36. 18.  The numerical features extracted from the training set were related, among others, to “2D Gaussian function,” “spatial compactness,” “contrast-based filtering,” “high-dimensional Gaussian filters,” and “element uniqueness.” In chapter 6, using the case of the “2D Gaussian function,” I will deal with these formulating practices.

37. 19.这可以理解为对最近不断增长的重要算法偏见文献的温和批评。Obermeyer 等人（2019 年）、Srivastava 和 Rossi（2018 年）以及 Yapo 和 Weiss（2018 年）等作者表明，许多算法的结果确实受到构建者的先入之见的影响。虽然我认为这种说法完全正确——算法源于受思维和行动习惯影响的问题化实践——但它也存在混淆前提与后果的风险：偏见不是算法的后果，但也许是使算法存在的因素之一。在我看来，某些由基本事实表达和实现的偏见可以而且应该被视为有害、不公正和错误的；例如，种族和性别偏见必须受到挑战和争议。然而，这些争议的结果很可能是其他可能危害较小、不公正和不正确的基本事实中表达的其他偏见。就算法而言，一种偏见会导致另一种偏见；因此，确认它们的存在并让它们可见非常重要，以便最终使它们与人们希望算法促进的价值观保持一致。

38. 19.  This can be read as a mild critique of the recent, growing, and important literature on algorithm biases. Authors such as Obermeyer et al. (2019), Srivastava and Rossi (2018), and Yapo and Weiss (2018), among others, show that the results of many algorithms are indeed biased by the preconceptions of those who built them. Though this statement is, I believe, completely correct—algorithms derive from problematization practices influenced by habits of thought and action—it also runs the risk of confusing premises with consequences: biases are not the consequences of algorithms but, perhaps, are one of the things that make them come into existence. Certain biases expressed and materialized by ground truths can and, in my opinion, should be considered harmful, unjust, and wrong; racial and gender biases have, for example, to be challenged and disputed. However, the outcome of these disputes may well be other biases expressed in other potentially less harmful, unjust, and incorrect ground truths. As far as algorithms are concerned, one bias calls for another; hence the importance of asserting their existence and making them visible in order to, eventually, align them with values one wishes algorithms to promote.

39. 20. Edwards (2013) 使用术语“数据图像”而不是“基本事实”。但我认为两者在某种程度上是等同的，指的是围绕数据组织的数字存储库，其值根据独立变量（尚需定义）而变化。

40. 20.  Edwards (2013) uses the term “data image” instead of “ground truth.” But I assume that both are somewhat equivalent and refer to digital repositories organized around data whose values vary according to independent variables (that yet need to be defined).

41. 21.在第 6 章的结尾，我将回到机器学习及其当代标签“人工智能”的主题。

42. 21.  At the end of chapter 6, I will come back to the topic of machine learning and its contemporary labeling as “artificial intelligence.”

43. 22.本讨论根据 2014 年 5 月至 10 月航海日志 3 中的笔记重建。

44. 22.  This discussion has been reconstructed from notes in Logbook 3, May–October 2014.

45. 23。然而，有趣的是，BJ 指责了图像处理领域重要会议的审稿人。据他所说，审稿人倾向于优先考虑那些做出“经典改进”的论文，而不是那些解决（从而定义）新问题的论文。无论如何，该小组的论文框架显然存在问题，因为审稿人并不相信其论点。因此，该算法无法在学术界和工业界传播，并且它的存在在一段时间内仅限于实验室的服务器。

46. 23.  However, it is interesting to note that BJ blames the reviewers of important conferences in image processing. According to him, the reviewers tend to privilege papers that make “classical improvement” over those that solve—and thus define—new problems. At any rate, there was obviously a problem in the framing of the Group’s paper as the reviewers were not convinced by its line of argument. As a consequence, the algorithm could not circulate within academic and industrial communities and its existence remained, for a while, circumscribed to the Lab’s servers.

笔记

Note

1. 1.在计算机科学和工程领域，人们确实普遍承认计算机编程实践很难进行，而且其结果非常不确定。关于这个有据可查的话题，请参阅 Knuth (2002)、Rosenberg (2008) 以及更文学化的 Ullman (2012a、2012b)。

2. 1.  In computer science and engineering, it is indeed well admitted that computer programming practices are difficult to conduct and their results very uncertain. On this well-documented topic, see Knuth (2002), Rosenberg (2008), and in a more literary way, Ullman (2012a, 2012b).

报告及其后果

A Report and Its Consequences

逐渐被称为“冯·诺依曼架构”的基石之一是约翰·冯·诺依曼于 1945 年匆忙撰写的EDVAC 报告初稿，该报告总结了宾夕法尼亚大学摩尔电气工程学院在二战期间发起的一项大胆电子计算系统的进展。我认为这份报告对搁置计算机系统的实际实例产生了重要影响，因此我们首先需要了解这份文件的历史和传播情况，以及它参与制定的世界。

One cornerstone of what will progressively be called “von Neumann architecture” is the First Draft of a Report on the EDVAC that John von Neumann wrote in a hurry in 1945 to summarize the advancement of an audacious electronic computing system initiated during World War II at the Moore School of Electrical Engineering at the University of Pennsylvania. As I believe this report has had an important influence on the setting aside of the practical instantiations of computer systems, we first need to look at the history and dissemination of this document as well as the world it participated in enacting.

编程心理学（及其局限性）

The Psychology of Programming (And Its Limits)

实践的问题在于必须做事：本质即存在，存在即行动（Deleuze 1995）。一旦电子计算系统开始被呈现为由中央器官控制的输入输出功能设备，使它们以期望的方式运行所需的努力很快就凸显出来：让这些设备做有意义的事情是极其繁琐的。这些智能电子大脑在实践中就像洗碗水一样乏味。但人们并没有对第一稿的输入输出框架产生怀疑，认为它在形式上很出色，但在经验上却不准确，而是很快将责任归咎于负责设计计算机输入的个人。简而言之，如果一个人无法让电子大脑运行，那是因为他没有设法给它们应得的输入。很快被称为“编程心理学”的学科试图理解为什么个人与电子计算机的交互如此费力。

The problem with practice is that it is necessary to do things: essence is existence and existence is action (Deleuze 1995). And as soon as electronic computing systems started to be presented as input-output functional devices controlled by a central organ, the efforts required to make them function in desired ways quickly stood out: it was extremely tedious to make the devices do meaningful things. These intelligent electronic brains were, in practice, dull as dishwater. But rather than casting doubts on the input-output framework of the First Draft and considering it formally brilliant but empirically inaccurate, the blame was soon casted on the individuals responsible for the design of computer’s inputs. In short, if one could not make electronic brains operate, it was because one did not manage to give them the inputs they deserved. What was soon called the “psychology of programming” tried, and tries, to understand why individuals interact so laboriously with electronic computers.

这种对个人的重视首先导致了20 世纪 50 年代能力倾向测试的出现，该测试旨在在劳动力稀缺的时代选出适合编程工作的候选人。到 20 世纪 70 年代末，西方软件行业从科学工艺转向性别工程的复杂动态推动了行为研究的开展，这些研究通常由编程测试组成，其相对结果归因于受控参数。十年后，这些行为测试的争议结果以及心理学学科内的理论争论导致了对编程的认知研究。认知科学家抛开了行为主义者提出的参数概念，专注于程序员应该开发的心理模型，以构建高效的程序。正如我们将看到的，这些研究工作以一种阻止他们探究程序员所做的事情的方式构建了编程，从而使他们的日常工作一直被隐藏起来。

This emphasis on the individual first led to aptitude tests in the 1950s that aimed at selecting the appropriate candidates for programming jobs in a time of workforce scarcity. By the late 1970s, entangled dynamics that made Western software industry shift from scientific craft to gender-connoted engineering supported the launching of behavioral studies that typically consisted of programming tests whose relative results were attributed to controlled parameters. A decade later, the contested results of these behavioral tests as well as theoretical debates within the discipline of psychology led to cognitive studies of programming. Cognitive scientists put aside the notion of parameters as proposed by behaviorists to focus on the mental models that programmers should develop to construct efficient programs. As we shall see, these research endeavors framed programming in ways that prevented them from inquiring into what programmers do, thus perpetuating the invisibilization of their day-to-day work.

人员选拔和能力测试

Personnel Selection and Aptitude Tests

到 20 世纪 40 年代末，在冯·诺依曼体系结构启发的第一批电子计算系统完成的同时，这些系统的实际处理问题也出现了：这些自动机似乎高度异质。这个实际问题很快在拥有第一台电子计算机的大学中出现。正如莫里斯·威尔克斯在关于 EDSAC 的回忆录中所写：

By the end of the 1940s, simultaneous to the completion of the first electronic computing systems that the von Neumann architecture inspired, the problem of the actual handling of these systems arose: these automatons appeared to be highly heteronomous. This practical issue quickly arose in the universities hosting the first electronic computers. As Maurice Wilkes wrote in his memoirs about the EDSAC:

到 1949 年 6 月，人们开始意识到，让程序正确运行并不像以前那么容易。我清楚地记得我第一次意识到这一点时的情景。EDSAC 位于大楼的顶层，而磁带打孔和编辑设备则位于楼下，位于一个环绕微分分析仪房间的走廊上。我当时正试图运行我的第一个非平凡程序，这是一个用于 Airy 微分方程的数值积分的程序。正是在一次往返于 EDSAC 房间和打孔设备之间的旅途中，“在楼梯拐角处犹豫不决”时，我突然意识到，我余生的大部分时间都将花在寻找自己程序中的错误上。（Wilkes 1985，145）

By June 1949 people had begun to realize that it was not so easy to get programs right as at one time appeared. I well remember when this realization first came on me with full force. The EDSAC was on the top floor of the building and the tape-punching and editing equipment one floor below on a gallery that ran round the room in which the differential analyzer was installed. I was trying to get working my first non-trivial program, which was one for the numerical integration of Airy’s differential equation. It was on one of my journeys between the EDSAC room and the punching equipment that “hesitating at the angles of stairs” the realization came over me with full force that a good part of the remainder of my life was going to be spent in finding errors in my own programs. (Wilkes 1985, 145)

尽管 EDSAC 理论上包含了所有可能的程序，但在具体情况下如何实现这些程序才是主要的实际问题。威尔克在直接尝试让这个功能设备发挥作用时，就意识到了这一点。

Although the EDSAC theoretically included all possible programs, the actualization of these programs within specific situations was the main practical issue. And this became obvious to Wilke once he was directly involved in trying to make the functional device function.

在工业界，电子计算系统的异质性方面也迅速崛起。第一个例子是围绕 UNIVAC（通用自动计算机的缩写）的争议，这是埃克特和莫奇利在 1946 年离开摩尔学院创办自己的公司（雷明顿兰德公司很快收购了这家公司）后开发的电子计算系统。当整个编程团队（由约翰·莫奇利领导）让 UNIVAC 运行一个统计程序，准确预测了 1952 年美国总统大选的结果时，UNIVAC 的潜力得到了广泛关注。这一营销举措的成本被小心地隐瞒了，进一步扩大了功能电子大脑接收输入并产生巧妙输出的形象。但当通用电气公司在 1954 年收购了一台 UNIVAC 计算机时，它很快意识到了系统展示与实际实施之间的差距：让这个功能系统发挥作用根本是不可能的。而且直到两年后，在聘请了一支全新的编程团队之后，一套基本的会计应用程序才开始产生一些有意义的结果结果（Campbell-Kelly 2003，25-30）。IBM 的 701 计算系统也面临类似的问题。平稳自动化的承诺很快就遇到了实践的现实：IBM 701 的首批用户——尤其是波音、通用汽车和国家安全局（Smith 1983）——不得不聘请专门致力于让系统发挥有用作用的整个团队^。17

In the industry, the heteronomous aspect of electronic computing systems also quickly stood up. A first example is the controversies surrounding the UNIVAC—an abbreviation for Universal Automatic Computer—an electronic computing system that Eckert and Mauchly developed after they left the Moore School in 1946 to launch their own company (which Remington Rand soon acquired). The potential of the UNIVAC gained a general audience when a whole programming team—which John Mauchly headed—made it run a statistical program that accurately predicted the results of 1952 American presidential election. This marketing move, whose costs were carefully unmentioned, further expanded the image of a functional electronic brain receiving inputs and producing clever outputs. But when General Electric acquired a UNIVAC computer in 1954, it quickly realized the gap between the presentation of the system and its actual enactment: it was simply impossible to make this functional system function. And it was only after two years and the hiring of a whole new programming team that a basic set of accounting applications could start producing some meaningful results (Campbell-Kelly 2003, 25–30). IBM faced similar problems with its computing system 701. The promises of smooth automation quickly faced the down-to-earth reality of practice: the first users of IBM 701—notably Boeing, General Motors, and the National Security Agency (Smith 1983)—had to hire whole teams specifically dedicated to making the system do useful things.¹⁷

美国国防机构也面临着同样的问题。1949 年 8 月苏联第一颗原子弹爆炸后，美国显得非常脆弱；现有的防空系统及其缓慢的手动雷达数据收集和处理方式根本无法及早发现核轰炸机，从而组织拦截机的反击行动。这种威胁以及许多其他远远超出本章范围的纠缠因素导致了能够实时处理雷达数据的原型计算机系统的开发。18^原型的良好结果进一步表明，1954 年应该实现一个全国性的高速数据处理系统防御系统，称为半自动地面环境(SAGE)。19^美国空军联系了许多承包商来工业化开发这个系统，IBM 被授予开发 250 吨重的 AN/FSQ-7 电子计算机的任务。20^但这些知名机构（包括 IBM、通用电气、贝尔实验室和麻省理工学院）都没有接受开发使这种强大的计算机可用的指令列表。几乎是理所当然，这份价值 2000 万美元的合同被授予了兰德公司，该公司是一家非营利（但非慈善）的政府组织，成立于 1948 年，是美国空军的一个研究部门。兰德公司之前曾参与过 SAGE 项目的开发，但其 25 名程序员组成的团队显然太小，无法胜任新的编程任务。因此，到 1956 年，兰德公司在全国范围内开始了一场重要的招聘活动，以寻找能够成功完成编程任务的人才。

US defense agencies were confronted with the same issue. After the explosion of the first Soviet atomic bomb in August 1949, the United States appeared dangerously vulnerable; the existing air defense system and its slow manual gathering and processing of radar data could by no means detect nuclear bombers early enough to organize counter operations of interceptor aircrafts. This threat—and many other entangled elements that are far beyond the scope of this chapter—led to the development of a prototype computer-based system capable of processing radar data in real time.¹⁸ The promising results of the prototype further suggested in 1954 the realization of a nationwide defense system of high-speed data-processing systems—called Semi-Automatic Ground Environment (SAGE).¹⁹ The US Air Force contacted many contractors to industrially develop this system of systems, with IBM being awarded the development of the 250 tons AN/FSQ-7 electronic computers.²⁰ But none of these renowned institutions—among them IBM, General Electric, Bell Labs, and MIT—accepted the development of the lists of instructions that would make such powerful computers usable. Almost by default, the $20 million contract was awarded to the RAND Corporation, a nonprofit (but nonphilanthropic) governmental organization created in 1948 that operated as a research division for the US Air Force. RAND had already been involved in the previous development of the SAGE project, but its team of twenty-five programmers was obviously far too small for the new programming task. So by 1956, RAND started an important recruiting campaign all around the country to find individuals who could successfully pursue the task of programming.

在冷战初期，兰德公司面临的挑战是在短时间内招募大量编程人员。为了满足这一大规模人员选拔的要求，兰德公司系统开发部的心理学家开始开发测试，其定量结果可以与未来的编程能力呈正相关。这些能力测试在很大程度上受到瑟斯顿初级智力测试²¹的启发，尽管兰德公司内部对此提出了批评（Rowan 1956），但很快就成为这是选拔新程序员的主要依据，因为它们可以节省大量时间，同时又以统计驱动的心理测量学为基础。大量使用能力测试帮助兰德公司迅速增加了程序员队伍，以至于其系统开发部门很快就被并入一个独立的组织，即系统开发公司(SDC)。早在 1959 年，SDC 就有“700 多名程序员在 SAGE 上工作，还有 1,400 多人为他们提供支持。……这估计占美国全部编程人力的一半”（Campbell-Kelly 2003，39）。但除了使兰德公司/SDC 能够更自信地参与 SAGE 项目外，能力测试还对编程工作的构思产生了重要影响。虽然这些测试的主要目标是支持快速和全国性的人员选拔，但它们也有助于将编程定义为一组抽象的智力操作，可以使用代理来衡量。

In this early Cold War period, the challenge for RAND was then to recruit a lot of programming staff in a short period of time. And to equip this massive personnel selection imperative, psychologists from RAND’s System Development Division started to develop tests whose quantitative results could positively correlate with future programming aptitudes. Largely inspired by the Thurstone Primary Mental Abilities Test,²¹ these aptitude tests—although criticized within RAND itself (Rowan 1956)—soon became the main basis for the selection of new programmers as they allowed crucial time savings while being based on the statistically driven discipline of psychometrics. The intensive use of aptitude tests helped RAND to rapidly increase its pool of programmers, so much so that its System Development Division was soon incorporated into a separate organization, the System Development Corporation (SDC). As early as 1959, the SDC had “more than 700 programmers working on SAGE, and more than 1,400 people supporting them. … This was reckoned to be half of the entire programming manpower of the United States” (Campbell-Kelly 2003, 39). But besides enabling RAND/SDC to engage more confidently in the SAGE project, aptitude tests also had an important effect on the very conception of programming work. Although the main goal of these tests was to support a quick and nationwide personnel selection, they also contributed to framing programming as a set of abstract intellectual operations that can be measured using proxies.

SDC 发起的能力测试制度迅速传遍整个行业，尤其​​促使 IBM 于 1959 年开发了自己的问卷，以支持其同样重要的招聘需求。IBM 编程能力测试 (PAT) 与从电子计算的开创性时期继承下来的计算机与大脑的相似性非常吻合，通常要求求职者找出表格之间的类比、继续数字列表并解决算术问题（见图3.1）。尽管求职者的能力测试分数与他们未来的工作表现之间的相关性仍存在争议，但能力测试很快成为 20 世纪 60 年代购买电子计算机的公司和行政部门的主流招聘工具。正如 Ensmenger (2012, 64) 所指出的那样：“到 1962 年，估计有 80% 的企业在招聘程序员时使用某种形式的能力测试，其中一半使用了 IBM PAT。”这些测试在新兴计算机行业中的大规模传播和使用进一步限制了编程实践作为可衡量的天生智力能力的范围。

The regime of aptitude testing as initiated by the SDC quickly spread throughout the industry, notably prompting IBM to develop its own questionnaire in 1959 to support its similarly important recruitment needs. Well in line with the computer-brain parallel inherited from the seminal period of electronic computing, the IBM Programming Aptitude Test (PAT) typically asked job candidates to figure out analogies between forms, continue lists of numbers, and solve arithmetic problems (see figure 3.1). Though the correlation between candidates’ scores to aptitude tests and their future work performances was a matter of debate, aptitude tests quickly became mainstream recruiting tools for companies and administrations that purchased electronic computers during the 1960s. As Ensmenger (2012, 64) noted: “By 1962, an estimated 80 percent of all businesses used some form of aptitude test when hiring programmers, and half of these used IBM PAT.” The massive distribution and use of these tests among the emerging computing industry further constricted the framing of programming practices as measurable innate intellectual abilities.

假定危机和行为研究

Supposed Crisis and Behavioral Studies

通过将编程定义为需要个人直觉素质的活动，能力测试在一定程度上抵消了与大学学位不平等相关的性别歧视。正如 Abbate (2012, 52) 指出的那样：“一个从未有机会获得大学学位的女性——或者被引导进入非技术专业——可以找到一份工作面试、参加考试，你就能立即获得未来程序员的资格。”与美国绝大多数技术性职业不同，计算机编程从一开始就有女性工作者参与，其中一些女性在战争期间就已经参与了计算项目。

By framing programming as an activity requiring personal intuitive qualities, aptitude tests have somewhat worked against gendered discriminations related to unequal access to university degrees. As Abbate (2012, 52) noted: “A woman who had never had the chance to earn a college degree—or who had been steered into a nontechnical major—could walk into a job interview, take a test, and instantly acquire credibility as a future programmer.” From its inception, computer programming, unlike the vast majority of skilled technical professions in the United States, has involved women workers, some of whom had already taken part to computing projects during the war.

然而，与 20 世纪 50 年代末的大多数西方职业环境一样，新兴的计算机行业受到普遍的刻板印象的推动，这种刻板印象往往阻碍女性程序员担任高级管理职位，并鼓励她们从事关系客户服务工作。这些性别动态不容忽视，因为它们有助于理解巧妙的软件设备快速且往往被低估的发展。由于女性在计算机相关专业领域中的独特地位——既是专家从业者，又往往是面向客户的代表——尽管她们在行业中所占比例相当小，但她们积极参与创新，旨在让专家和新手都更容易编程。最著名的例子当然是 UNIVAC 的编程主管 Grace Murray Hopper，她于 1951 年开发了第一个编译器（一种将其他程序转换为机器代码的程序），^之后于 1955 年设计了商业编程语言 B-0（更名为 FLOW-MATIC）。但在整个 20 世纪 50 年代和 60 年代，还有许多其他女性积极参与软件创新，尽管她们往往活在更为显眼的男性经理的阴影下。这些重要人物包括 20 世纪 50 年代中期开发了广泛使用的数据编辑代码的 Adele Mildred Koss 和 Nora Moser；负责 FORTRAN 高级编程语言流程分析的 Lois Haibt；以及 20 世纪 50 年代后期通用商业语言 (COBOL) 项目先锋 Mary Hawes、Jean Sammet 和 Gertrude Tierney（Abbate 2012, 79–81）。

However, like most Western professional environments in the late 1950s, the nascent computing industry was fueled by pervasive stereotypes, often preventing women programmers from occupying upper managerial positions and encouraging them to do relational customer care work. These gender dynamics should not be overlooked as they help to understand the rapid, and often underappreciated, development of ingenious software equipment. Due to their unique position within the computer-related professional worlds—both expert practitioners and, often, representatives toward clients—women, given their rather small percentage within the industry, actively contributed to innovations aimed at making programming easier for experts and novices alike. The most notorious example is certainly Grace Murray Hopper, head of programming for UNIVAC, who developed the first compiler—a program that translates other programs into machine code²²—in 1951 before designing the business programming language B-0 (renamed FLOW-MATIC) in 1955. But many other women actively took part to software innovations throughout the 1950s and 1960s, though often in the shadow of more visible male managers. Among these important figures are Adele Mildred Koss and Nora Moser who developed widely used code for data editing in the mid-1950s; Lois Haibt who was responsible for flow analysis of the FORTRAN high-level programming language; and Mary Hawes, Jean Sammet, and Gertrude Tierney who were at the forefront of the common business-oriented language (COBOL) project in the late 1950s (Abbate 2012, 79–81).

从 20 世纪 60 年代中期开始，对编译器和高级编程语言的改进（这些改进通常来自女性）也使计算能力增加了 10 倍（Mody 2017，47–77）。这种新的有前途的软件和硬件基础设施的结合促使大型标志性计算机制造商开始构建越来越复杂的程序，例如操作系统和大型商业应用程序。其中一些备受瞩目的项目（如 IBM 的 System 360 项目）的彻底失败，²³很快引起了当时评论家的不确定性，其中一些评论家使用了令人回味的计算机历史学家对这场软件危机的真实性表示怀疑，因为精确的调查显示，除了一些引人注目的非标准项目外，20 世纪 60 年代末的软件生产总体上是按时按预算进行的 (Campbell-Kelly 2003, 94)。但危机言论（也以夸张但流行的软件生产成本话语为依据）²⁴仍然对该行业产生了切实的影响，甚至改变了其整体方向和身份。

From the mid-1960s onward, refinements over compilers and high-level programming languages, which had often come from women, were added to the impressive tenfold increase in computing power (Mody 2017, 47–77). This combination of new promising software and hardware infrastructures prompted large iconic computer manufacturers to start building increasingly complex programs, such as operating systems and massive business applications. The resounding failures of some of these highly visible projects, like the IBM project System 360,²³ soon gave rise to a sense of uncertainty among commentators at the time, some of whom used the evocative expression of “software crisis” (Naur and Randell 1969, 70–73). Historians of computing have expressed doubts about the reality of this software crisis as precise inquiries have shown that, apart from some highly visible and nonstandard projects, software production in the late 1960s was generally on time and on budget (Campbell-Kelly 2003, 94). But the crisis rhetoric, which also fed on an exaggerated but popular discourse on software production costs,²⁴ nonetheless had tangible effects on the industry to the point of changing its overall direction and identity.

与相关的微电子学科相比，编程长期以来缺乏可信度和声望。尽管在 20 世纪 50 年代和 60 年代取得了重大进展，但参与软件生产的参与者在西方计算研究和行业中的地位往往较低。对于女性程序员来说，情况确实如此，因为她们在技术环境中工作。但对于男性程序员来说也是如此，因为他们在包括女性的领域工作。从这个角度来看，20 世纪 60 年代末盛行的危机言论——以不代表行业现状的标志性失败为食——提供了一个机会，可以根据当时的标准将编程重新定义为更有价值的东西（Ensmenger 2010，195-222）。这可能是积极意义的“工程”一词开始传播并成为一种视线的原因之一，特别是通过 1968 年北大西洋公约组织 (NATO) 题为“软件工程”的会议的努力以及专业组织和学术期刊的建立，例如电气和电子工程师协会的《IEEE 软件工程学报》（1975 年）和计算机协会的《ACM 软件工程笔记》（1976 年）。尽管一些知名人士认为软件生产已经是严格和系统的，但他们对这一复杂的学科重新贴标签的过程表示反对，许多程序员（包括女性和男性）都支持这一过程，他们认为工程师的头衔是改善工作条件的机会。然而，正如 Abbate (2012, 104) 指出的那样：“这一举措的一个意想不到的后果可能是让编程和计算机科学对女性的吸引力降低，这有助于解释为什么女性在第一波软件改进中发挥了主导作用，但在软件工程时代却变得不那么引人注目的历史谜团。”

When compared with the related discipline of microelectronics, programming has long suffered from a lack of credibility and prestige. Despite significant advances throughout the 1950s and the 1960s, actors taking part to software production were often accorded a lower status within Western computing research and industry. This was true for women programmers since they were working in a technical environment. But it was also true for men programmers since they were working in a field that included women. Under this lens, the crisis rhetoric that took hold at the end of the 1960s—feeding on iconic failures that were not representative of the state of the industry—provided an opportunity to reinvent programming as something more valuable according to the criteria of the time (Ensmenger 2010, 195–222). This may be one of the reasons why the positively connoted term “engineering” started to spread and operate as a line of sight, notably via the efforts of the 1968 North Atlantic Treaty Organization (NATO) conferences entitled “Software Engineering” and the setting up of professional organizations and academic journals such as the Institute of Electrical and Electronics Engineers’ IEEE Transactions on Software Engineering (1975) and the Association for Computing Machinery’s ACM Software Engineering Notes (1976). Though contested by eminent figures who considered that software production was already rigorous and systematic, this complex process of disciplinary relabeling was supported by many programmers—women and men—who saw the title of engineer as an opportunity to improve their work conditions. However, as Abbate (2012, 104) pointed out: “An unintended consequence of this move may have been to make programming and computer science less inviting to women, helping to explain the historical puzzle of why women took a leading role in the first wave of software improvements but become much less visible in the software engineering era.”

这种让软件生产走上工程化道路的愿望被认为是解决所谓危机的解决方案，而这种危机本身就建立在性别低估编程工作——这一现象已经影响到了编程的学术分析。与这一学科重新定位相平行的是，自称行为主义传统的实证主义研究在 20 世纪 70 年代初开始对编程工作产生兴趣。对于这些研究人员来说，分析重点应该转移：学者们不应该定义编程和设计能力测试所需的固有技能，而应该尝试提取导致最佳编程性能的参数并提出改进软件生产的方法。高级编程语言的引入和传播以及计算机科学学术课程的增加在很大程度上参与了这一新研究方向的建立。有了 FORTRAN 或 COBOL 等不依赖于计算机特性和品牌的编程语言，行为心理学家和计算机科学家能够在受控环境中设计编程测试。此外，计算机科学学术课程的增加提供了相对多样化的人群（例如本科生、研究生、教职员工），他们可以通过这些编程测试。这两个要素使得实验设计成为可能，该实验根据不同参数集（年龄、经验、设计辅助工具）可能产生的结果对其进行排序（见图3.2）。

This stated desire to make software production take the path of engineering—considered the solution to a supposed crisis that itself built on a gendered undervaluation of programming work—has rubbed off on the academic analysis of programming. Parallel to this disciplinary reorientation, a line of positivist research claiming behaviorist tradition began to take an interest in programming work in the early 1970s. For these researchers, the analytical focus should shift: instead of defining the inherent skills required for programming and design aptitude tests, scholars should rather try to extract the parameters that induce the best programming performances and propose ways to improve software production. The introduction and dissemination of high-level programming languages as well as the multiplication of academic curricula in computer science highly participated in establishing this new line of inquiry. With programming languages such as FORTRAN or COBOL that did not depend on the specificities and brands of computers, behavioral psychologists along with computer scientists became able to design programming tests in controlled environments. Moreover, the multiplication of academic curricula in computer science provided relatively diverse populations (e.g., undergrads, graduates, faculty members) that could pass these programming tests. These two elements made possible the design of experiments that ranked different sets of parameters (age, experience, design aids) according to the results they assumedly produced (see figure 3.2).

该框架引发了大量关于调试性能的测试（例如，Bloom 1980；Denelesky and McKee 1974；Sackman、Erikson 和 Grant 1968；Weinberg 1971，122-189；Wolfe 1971），设计辅助性能（例如，Blaiwes 1974；Brooke and Duncan，1980a，1980b；Kammann 1975；Mayer 1976；Shneiderman 等人 1977；Weinberg 1971，205-281；Wright and Reid 1973）和逻辑语句性能²⁵（例如，Dunsmore and Gannon 1979；Gannon 1976；Green 1977；Lucas and Kaplan 1976；Sime、Green 和 Guest 1973；Sime、Arblaster 和 Green 1977；Sime、Green 和 Guest 1977；Sheppard 等人 1979；Weissman 1974）。但是，尽管这些研究具有系统性，但它们的结果却显而易见，正如 Curtis (1988) 所解释的那样，软件承包商无需正式参与行为实验，就已经知道，例如，经验丰富的程序员比没有经验的程序员能产生更好的结果，或者流程图或文档等设计辅助工具是编程实践的有用工具。这些一般和冗余的事实并不能帮助程序员更好地设计指令列表。到 20 世纪 80 年代，越来越强大的计算系统仍然非常不稳定。操作起来很困难，尽管它们是由在越来越男性化的环境中工作的软件工程师指导的。

This framework led to numerous tests on debugging performances (e.g., Bloom 1980; Denelesky and McKee 1974; Sackman, Erikson, and Grant 1968; Weinberg 1971, 122–189; Wolfe 1971), design aid performances (e.g., Blaiwes 1974; Brooke and Duncan, 1980a, 1980b; Kammann 1975; Mayer 1976; Shneiderman et al. 1977; Weinberg 1971, 205–281; Wright and Reid 1973), and logical statement performances²⁵ (e.g., Dunsmore and Gannon 1979; Gannon 1976; Green 1977; Lucas and Kaplan 1976; Sime, Green, and Guest 1973; Sime, Arblaster, and Green 1977; Sime, Green, and Guest 1977; Sheppard et al. 1979; Weissman 1974). But despite their systematic aspect, these studies suffered from the obviousness of their results, for as explained by Curtis (1988), without formally being engaged in behavioral experiments, software contractors were already aware that, for example, experienced programmers produced better results than inexperienced ones did, or that design aids such as flowcharts or documentation were helpful tools for the practice of programming. These general and redundant facts did not help programmers to better design lists of instructions. By the 1980s, the increasingly powerful computing systems remained terribly difficult to operate, be they instructed by software engineers working in more and more malely connoted environments.

认知转向

The Cognitive Turn

到 20 世纪 70 年代末，行为主义观点开始受到心理学界的批评。对于越来越多的认知心理学家（有时在人工智能部门工作）来说，行为研究结果的明显性似乎是方法论缺陷的结果，许多经过排序的参数集收集了重要的个体结果差异。据几位认知研究人员称，行为研究的分析单位是错误的；由于许多结果的差异存在于同一组参数中，因此对这些参数集进行排序根本毫无意义（Brooks 1977、1980；Curtis 1981；Curtis 等人 1989；Moher 和 Schneider 1981）。这些认知主义者提出的解释他们所谓的“个体差异”的解决方案是深入研究以便更好地理解计算机程序形成背后的认知过程和心理模型。

By the end of the 1970s, the behavioral standpoint began to be criticized from inside the psychological field. To more and more cognitive psychologists, sometimes working in artificial intelligence departments, it seemed that the obviousness of behavioral studies’ results was function of a methodological flaw, with many of the ranked sets of parameters gathering important individual variations of results. According to several cognitive researchers, the unit of analysis of behavioral studies was erroneous; since many results’ disparities existed within the same sets of parameters, the ranking of these sets was simply senseless (Brooks 1977, 1980; Curtis 1981; Curtis et al. 1989; Moher and Schneider 1981). The solution that these cognitivists proposed to account for what they called “individual differences” was then to dive inside the individuals’ head to better understand the cognitive processes and mental models underlying the formation of computer programs.

“程序”和“认知”概念之间的紧密联系也使计算机编程研究对认知科学家具有吸引力。正如 Ormerod (1990, 63–64) 所说：

The strong relationships between the notions of “program” and “cognition” also participated in making the study of computer programming attractive to cognitive scientists. As Ormerod (1990, 63–64) put it:

认知和编程领域主要有三种关联。首先，认知心理学基于“计算隐喻”，其中人类思维被视为一种类似于计算机的信息处理器。其次，认知心理学提供了检查计算任务性能背后过程的方法。第三，编程是一项定义明确的任务，程序员的数量正在增加，这使其成为研究现实世界领域认知过程的理想任务。

The fields of cognition and programming are related in three main ways. First, cognitive psychology is based on a “computational metaphor,” in which the human mind is seen as a kind of information processor similar to a computer. Secondly, cognitive psychology offers methods for examining the processes underlying performance in computing tasks. Thirdly, programming is a well-defined task, and there are an increasing number of programmers, which makes it an ideal task in which to study cognitive process in a real-world domain.

这三个因素——认知和计算机程序之间假定的基本相似性、程序员人数的不断增长以及可用于研究这一群体的可用方法——极大地促使认知科学家将计算机编程视为一个富有成果的研究课题。此外，投资行为主义者未能理解的话题也被视为展示认知主义方法优越性的机会。在一定程度上，其目的也是为了表明行为是心理过程的函数：

These three elements—the assumed-fundamental similarity between cognition and computer programs, the growing population of programmers, and the available methods that could be used to study this population—greatly contributed to making cognitive scientists consider computer programming as a fruitful topic of inquiry. Moreover, investing in a topic that behaviorists failed to understand was also seen as an opportunity to demonstrate the superiority of cognitivist approaches. To a certain extent, the aim was also to show that behaviors were a function of mental processes:

[行为主义者] 试图建立描述编程行为的各种参数的有效性，而不是试图指定决定这些参数的底层过程。（Brooks 1977，740）

[Behaviorists] attempt to establish the validity of various parameters for describing programming behavior, rather than attempting to specify underlining processes which determine these parameters. (Brooks 1977, 740)

当时的目标是描述导致良好编程表现的心理过程，并最终利用这些心理过程来培训或选拔更好的程序员。然而，认知研究的方法大多数时候与编程行为研究的方法并没有根本区别。针对不同的个体，通常是计算机科学专业的学生或教职员工，提出了特定的编程测试。然后根据测试的正确答案以及受计算思维隐喻启发的人类理解的一般认知模型（尤其是 Newell 和 Simon [1972] 以及后来的 Anderson [1983] 的模型），分析个体的回答、评论（口头或书面）和元数据（按键次数、花在问题上的时间等）。从结果、评论和一般认知模型之间的这种对抗中，不同的心理模型具体对计算机编程任务的执行情况进行了推断、分类和排名，并根据其表现进行了排名（见图3.3）。

The ambition was then to describe the mental processes that lead to good programming performances and eventually use these mental processes to train or select better programmers. The methodology of cognitive studies was, most of the time, not radically different from that of behavioral studies on programming, though. Specific programming tests were proposed to different individuals, often computer science students or faculty members. The responses, comments (oral or written), and metadata (number of key strokes, time spent on the problem, etc.) of the individuals were then analyzed according to the rights answers of the test as well as based on general cognitive models of human understanding that the computational metaphor of the mind has inspired (especially the models of Newell and Simon [1972] and, later, Anderson [1983]). From this confrontation among results, comments, and general models of cognition, different mental models specific to the task of computer programming were inferred, classified, and ranked according to their performances (see figure 3.3).

这种计算机编程研究模式引发了大量研究，这些研究提出了解决抽象问题的心理模型（例如，Adelson 1981；Brooks 1977；Carroll、Thomas 和 Malhotra 1980；Jeffries 等人 1981；Pennington 1987；Shneiderman 和 Mayer 1979）和培养编程能力（例如，Barfield 1986；Coombs、Gibson 和 Alty 1982；McKeithen 等人 1981；Soloway 1986；Vessey 1989；Wiedenbeck 1985）。部分由于其结果有所缓和——正如 Draper (1992) 所承认的，认知主义者提出的众多心理模型并未显著提高编程性能——认知研究后来重新整合了行为主义的考虑（例如，受控的参数集），以获得今天的混合和以管理为中心的形式（Capretz 2014；Ahmed、Capretz 和 Campbell 2012；Ahmed 等人 2012；Cruz、da Silva 和 Capretz 2015）。

This research pattern on computer programming led to numerous studies proposing mental models for solving abstract problems (e.g., Adelson 1981; Brooks 1977; Carroll, Thomas, and Malhotra 1980; Jeffries et al. 1981; Pennington 1987; Shneiderman and Mayer 1979) and developing programming competencies (e.g., Barfield 1986; Coombs, Gibson, and Alty 1982; McKeithen et al. 1981; Soloway 1986; Vessey 1989; Wiedenbeck 1985). Due, in part, to their mitigated results—as admitted by Draper (1992), the numerous mental models proposed by cognitivists did not significantly contribute to better programming performances—cognitive studies have later reintegrated behaviorist considerations (e.g., controlled sets of parameters) to acquire the hybrid and management-centered form they have today (Capretz 2014; Ahmed, Capretz, and Campbell 2012; Ahmed et al. 2012; Cruz, da Silva, and Capretz 2015).

让认知回归本位

Putting Cognition Back to Its Place

大多数学术界试图更好地理解计算机编程的尝试似乎都存在令人讨厌的缺陷：能力倾向测试混淆了前提和后果，行为研究将行为置于黑匣子中，认知研究则陷入了同义反复的解释。如果我们要考虑计算机编程作为负责任的做法，我们似乎需要与这些勇敢但有问题的努力保持距离。

Most academic attempts to better understand computer programming seem to have annoying flaws: aptitude tests mix up premises and consequences, behavioral studies put actions into black boxes, and cognitive studies are stuck in tautological explanations. If we want to consider computer programming as accountable practices, it seems that we need to distance ourselves from these brave but problematic endeavors.

然而，即使我们的批评者是相关的，我们目前仍然无法提出任何替代方案。程序员的行为难道没有认知方面吗？程序员难道不用他们的头脑来计算解决复杂问题吗？认知和计算机程序之间的混淆很可能源于计算机历史的误导——正如我试图暗示的那样——它确立为普遍习惯的能力值得尊重。我们怎么能不把对计算机编程实践的实证研究当作愚蠢的简化呢？我们如何证明解释计算机编程行动过程的愿望是合理的，从而使这些行动过程可见，这些实践是任何计算机化项目的必经之路？

Yet, provided that our critics are relevant, we are at this point still unable to propose any alternative. Do the actions of programmers not have a cognitive aspect? Do programmers not use their minds to computationally solve complex problems? The confusion between cognition and computer programs may well derive from a misleading history of computers—as I tried to suggest—its capacity to establish itself as a generalized habit commands respect. How can we not present empirical studies of computer programming practices as silly reductions? How can we justify the desire to account for, and thus make visible, the courses of action of computer programming, these practices that are obligatory passage points of any computerization project?

幸运的是，当代哲学研究已经成功填补了认知与实践、智慧与愚钝行为之间的鸿沟。正是由于这些鼓舞人心的研究，我们才能够将编程视为一种实践，而不会完全抛弃认知的概念。为此，我首先需要快速重新考虑计算机是按照人类大脑和思维的形象设计的这一观点。正如我们已经看到的——尽管只是部分——这一观点只有在回顾时才有意义：具体发生的事情要复杂得多。然后，我将重新考虑将认知作为一个计算过程的哲学框架。最后，根据感知哲学的当代著作，我将研究一种认知定义，该定义保留了我们如何理解周围事物的重要方面，同时将其与实践和行动重新联系起来。通过假设主体在认知过程中的核心地位，这种认知的实施概念将进一步帮助我们从经验上考虑计算机编程过程中发生的事情。

Fortunately, contemporary work in philosophy has managed to fill in the gap that has separated cognition from practices, intelligent minds from dull actions. It is thanks to these inspiring studies that we will become able to consider programming as a practice without totally turning our back on the notion of cognition. To do so, I will first need to quickly reconsider the idea that computers were designed in the image of the human brain and mind. As we already saw—though partially—this idea is relevant only in retrospect: what has concretely happened is far more intricate. I will then reconsider the philosophical frame that encloses cognition as a computational process. Finally, following contemporary works in the philosophy of perception, I will examine a definition of cognition that preserves important aspects of how we make sense of the things that surround us while reconnecting it to practices and actions. By positing the centrality of agency in cognitive processes, this enactive conception of cognition will further help us empirically consider what is happening during computer programming episodes.

笔记

Notes

1. 1.我的出发点是任意的，因为我可以从其他地方、不同的时间开始。事实上，正如 Lévy ( 1995) 所指出的，所谓的“电子计算机的冯·诺依曼体系结构”的前提不仅可以在阿兰·图灵 1937 年的论文中找到，而且可以在 20 世纪 20 年代办公机器行业的发展中找到，也可以在查尔斯·巴贝奇 19 世纪下半叶的机械数学著作中找到，也可以在 18 世纪用穿孔卡编程的织布机中找到，等等，至少在莱布尼茨的二进制算术和帕斯卡的计算机之前都是如此。计算机的历史是模糊的。由于它只是“在对异质材料和设备进行一系列的改动和重新解释之后”才出现的（Lévy 1995，636），因此很难——事实上，几乎不可能——提出任何不纠缠的关联。幸运的是，本节的目的不是提供任何计算机历史：它“只是”试图提供一些元素，在我看来，这些元素参与形成了一个特定且有影响力的文档：冯·诺依曼关于 EDVAC 的报告。

2. 1.  My point of departure is arbitrary in the sense that I could have started somewhere else, at a different time. Indeed, as Lévy (1995) showed, the premises of what will be called “von Neumann architecture of electronic computers” can be found not only in Alan Turing’s 1937 paper but also in the development of the office-machine industry during the 1920s, but also in the mechanic-mathematical works of Charles Babbage during the second half of the nineteenth century, but also in eighteenth century’s looms programmed with punched cards, and so on, at least until Leibniz’s work on binary arithmetic and Pascal’s calculating machine. The history of the computer is fuzzy. As it only appears “after a cascade of diversions and reinterpretations of heterogeneous materials and devices” (Lévy 1995, 636), it is extremely difficult—in fact, almost impossible—to propose any unentangled filiation. Fortunately, this section does not aim to provide any history of the computer: It “just” tries to provide elements that, in my view, participated in the formation of one specific and influential document: von Neumann’s report on the EDVAC.

3. 2.有关第二次世界大战期间美国射击表设计的更精确说明，请参阅 Haigh、Priestley 和 Rope（2016 年，第 20-23 页）以及 Polachek（1997 年）。

4. 2.  For a more precise account of the design of firing tables in the United States during World War II, see Haigh, Priestley, and Rope (2016, 20–23) and Polachek (1997).

5. 3。除了有效的计算能力（需要长达数天的时间才能设置完成（Haigh、Priestley 和 Rope 2016，26），而且其结果通常不如手工计算准确（Polachek 1997，25–27））之外，微分分析仪的一个重要特征是它们能够吸引周围的计算专家。例如，到 1940 年，麻省理工学院、宾夕法尼亚大学和英国曼彻斯特大学（未来电子计算发展的三所重要机构）都拥有微分分析仪（Campbell-Kelly 等人 2013，45–50；Owens 1986）。关于微分分析仪在美国早期计算研究中的作用，另请参阅 Akera（2008，38–45）。

6. 3.  More than their effective computing capabilities—they required up to several days to be set up (Haigh, Priestley, and Rope 2016, 26) and their results were often less accurate than those provided by hand calculations (Polachek 1997, 25–27)—an important characteristic of differential analyzers was their capacities to attract computing experts around them. For example, by 1940, MIT, the University of Pennsylvania, and the University of Manchester, England—three important institutions for the future development of electronic computing—all possessed a differential analyzer (Campbell-Kelly et al. 2013, 45–50; Owens 1986). On the role of differential analyzers in early US-based computing research, see also Akera (2008, 38–45).

7. 4.影响射弹的众多因素的汇总始于阿伯丁的试验场，在那里测量了新设计的炮弹的速度（Haigh、Priestley 和 Rope 2016，20）。

8. 4.  The assembling of the numerous factors affecting the projectiles started at the test range in Aberdeen where the velocities of the newly designed shells were measured (Haigh, Priestley, and Rope 2016, 20).

9. 5.虽然定义炮弹轨迹计算的微分方程在数学上相当简单，但求解它们却非常复杂，因为需要对以非线性方式变化的空气阻力进行建模。正如 Haigh、Priestley 和 Rope (2016, 23) 所说：“与只选择响应优雅方法的方程的微积分老师不同，BRL 的数学家不能忽略风阻或分配不同的问题。与科学家和工程师制定的大多数微分方程一样，弹道方程需要更复杂的数值近似技术。”

10. 5.  Although the differential equations defining the calculation of shells’ trajectories are mathematically quite simple, solving them can be very complicated as one needs to model air resistance varying in a nonlinear manner. As Haigh, Priestley, and Rope (2016, 23) put it: “Unlike a calculus teacher, who selects only equations that respond to elegant methods, the mathematicians at the BRL couldn’t ignore wind resistance or assign a different problem. Like most differential equations formulated by scientists and engineers, ballistic equations require messier techniques of numerical approximation.”

11. 6 . 有趣的是，延迟线存储最初与雷达技术有关。更准确地说，1942 年雷达技术的一个问题是阴极射线管显示器显示移动和静止的物体。因此，雷达屏幕将飞机、建筑物或森林的位置转换为同一幅杂乱无章的画面，极难阅读。麻省理工学院的辐射实验室将移动目标指示器 (MTI) 的开发分包给摩尔学院，以开发一种可以根据雷达信号变化位置进行过滤的系统。这是摩尔学院延迟线存储技术的开端，最初与计算无关 (Akera 2008, 84–86; Campbell-Kelly et al. 2013, 69–74)。雷达技术还极大地帮助了 1943-1944 年英国高度机密的巨像计算机的设计 (L é vy 1995, 646)。

12. 6.  Interesting to note that delay-line storage is originally linked to radar technology. More precisely, one problem of the radar technology in 1942 was that cathode-ray tube displays showed moving and stationary objects. Consequently, radar screens translated the positions of planes, buildings, or forests in one same messy picture extremely difficult to read. MIT’s radiation laboratory subcontracted the development of a moving target indicator (MTI) to the Moore School in order to develop a system that could filter radar signals according to their changing positions. This was the beginning of delay-line storage technology at the Moore School, that at first had nothing to do with computing (Akera 2008, 84–86; Campbell-Kelly et al. 2013, 69–74). Radar technology also significantly helped the design of British highly confidential Colossus computer in 1943–1944 (Lévy 1995, 646).

13. 7 . 到 1942 年，为了加快弹道微分方程的求解速度，BRL 的人工计算机往往只考虑有限的因素。通过简化方程，可以生成和分发更多的射击表，但缺点是它们的精度往往会下降（Polachek 1997）。当然，在战争前线，一旦士兵意识到第一轮射击没有得到充分定义，他们仍然可以稍微修改远程武器的参数以提高其精度。然而——这是关键点——在第一轮射击和后续射击之间，对方有足够的时间进行掩护，从而使整个远程射击行动效率降低。战争的神经恰恰是第一次远距离齐射，如果准确的话，可以造成大量伤亡。因此，从某种程度上讲，战争的神经也是一种能力，可以在微分方程中加入更多因素，而微分方程的解会打印在射击表小册子中 (Haigh、Priestley 和 Rope 2016，25)。

14. 7.  By 1942, in order to speed up the resolution of ballistic differential equations, only a limited range of factors tended to be considered by human computers at the BRL. By simplifying the equations, more firing tables could be produced and distributed, but the drawback was that their precision tended to decrease (Polachek 1997). Of course, on the war front, once soldiers realized that the first volley was not adequately defined, they could still slightly modify the parameters of the long-distance weapon to increase its precision. Yet—and this is the crucial point—in between the first volley and the subsequent ones, the opposite side had enough time to take cover, hence making the overall long-distance shooting enterprise less effective. The nerve of war was precisely the first long-distance volleys that, when accurate, could lead to many casualties. By extension, then, the nerve of war was also, to a certain extent, the ability to include more factors in the differential equations whose solutions were printed out in firing table booklets (Haigh, Priestley, and Rope 2016, 25).

15. 8.国防研究委员会 (NDRC) 成立于 1940 年，将美国海军和战争部的研究实验室与数百所美国大学的实验室联合起来。NDRC 最初拥有一笔可观的预算，用于资助可能在未来战场上提供重大优势的应用研究项目。它还充当咨询机构，就像 ENIAC 的情况一样，由于需要大量不可靠的真空管，ENIAC 被认为几乎不可行。关于此主题，请参阅 Campbell-Kelly 等人 (2013, 70–72)。

16. 8.  Created in 1940, the National Defense Research Committee (NDRC) united the research laboratories of the US Navy and the Department of War with hundreds of US universities’ laboratory. The NDRC initially had an important budget to fund applied research projects that could provide significant advantages on future battlefields. It also operated as an advisory organization as in the case of the ENIAC that was considered nearly infeasible due to the important amount of unreliable vacuum tubes it would require. On this topic, see Campbell-Kelly et al. (2013, 70–72).

17. 9.这份合同的历史可以成为整本书的主题。有关其最重要时刻的精彩介绍，请参阅 Haigh、Priestley 和 Rope (2016, 17–33)。

18. 9.  The history of this contract could be the topic of a whole book. For a nice presentation of its most important moments, see Haigh, Priestley, and Rope (2016, 17–33).

19. 10 . Harvard Mark 1 是基于 Howard Aiken 的提议，由 IBM 于 1937 年至 1943 年底为哈佛大学开发的。尽管即使以当时的标准来看，它的计算速度也很慢，但它是一个重要的计算系统，因为它代表了科学计算和办公机器技术的早期融合。有关 Harvard Mark 1 的更深入历史，请参阅 Cohen (1999)。

20. 10.  Based on a proposal by Howard Aiken, the Harvard Mark 1 was developed by IBM for Harvard University between 1937 and late 1943. Though computationally slow, even for the standards of the time, it was an important computing system as it expressed an early convergence of scientific calculation and office-machine technologies. For a more in-depth history of the Harvard Mark 1, see Cohen (1999).

21. 11.尽管 ENIAC 在整个存在过程中其形状变化很大，但它本质上是一个由不同单元（累加器、乘法器和功能表）组成的网络。每个单元都有内置的刻度盘和开关。如果配置得当，这些刻度盘和开关可以定义一个操作；例如，“清除累加器的值”、“将数字传输给乘数 3”、“接收数字”等等。要开始处理操作，每个刻度盘和开关配置都必须由直接连接到特定单元的“程序线”触发。所有这些“程序线”形成了一个线路网络，将所有单元连接起来以执行一系列特定的操作。但是，一旦需要另一系列操作，就必须重新排列线路网络以适应新的刻度盘和开关配置。有关 ENIAC 设置的更多元素，请参阅 Haigh、Priestley 和 Rope 2016（35–57）。

22. 11.  Though its shape varied significantly throughout its existence, the ENIAC was fundamentally a network of different units (accumulators, multipliers, and function tables). Each unit had built-in dials and switches. If adequately configured, these dials and switches could define one single operation; for example, “clear the values of the accumulator,” “transmit a number to multiplier number 3,” “receive a number,” and so on. To start processing an operation, each configuration of dials and switches had to be triggered by a “program line” wired directly to the specific unit. All these “program lines” formed a network of wires connecting all the units for one specific series of operations. But as soon as another series of operations was required, the network of wires had to be rearranged in order to fit the new configurations of dials and switches. For more elements about the setup of ENIAC, see Haigh, Priestley, and Rope 2016 (35–57).

23. 12.冯·诺依曼曾想聘请艾伦·图灵担任普林斯顿大学的博士后助理。图灵拒绝了，因为他想回到英国 (MacRae 1999, 187–202)。

24. 12.  Von Neumann tried to hire Alan Turing as a postdoctoral assistant at Princeton. Turing refused as he wanted to return to England (MacRae 1999, 187–202).

25. 13.当然，曼哈顿计划是高度机密的，这使得冯·诺依曼无法向 ENIAC 团队说明他的计算需求。

26. 13.  The Manhattan Project was, of course, highly confidential and this prevented von Neumann from specifying his computational needs with the ENIAC team.

27. 14.正如 Akera (2008, 119–120) 和 Swade (2011) 所指出的，并由 Haigh、Priestley 和 Rope (2014; 2016, 231–257) 进一步论证的，“存储程序”的概念是一个历史产物：“‘存储程序概念’从未在公认的来源《第一稿》中作为一个特定特征提出，而只是追溯性地用来挑选出 EDVAC 设计的某些特征”(Haigh、Priestley 和 Rope 2016, 256)。

28. 14.  As suggested by Akera (2008, 119–120) and Swade (2011), and further demonstrated by Haigh, Priestley, and Rope (2014; 2016, 231–257), the notion of “stored program” is a historical artifact: “the ‘stored program concept’ was never proposed as a specific feature in the agreed source, the First Draft, and was only retroactively adopted to pick out certain features of the EDVAC design” (Haigh, Priestley, and Rope 2016, 256).

29. 15 。冯·诺依曼的《初稿》分发后不久，埃克特和莫奇利分发了一份更长、也远没那么出名的反报告，题为《自动高速计算：EDVAC 的进展报告》（Eckert and Mauchly 1945），在报告中他们强调了《初稿》的理想化方面。对于埃克特和莫奇利来说，风险确实很高：如果冯·诺依曼对 EDVAC 的理想化描述被认为是对该工程项目的真实描述，那么就永远无法从中获得专利。而事实也确实如此。1947 年，军械部的律师认定《初稿》是关于 EDVAC 项目的第一份出版物，因此取消了埃克特和莫奇利在 1946 年初提交的专利（Haigh、Priestley 和 Rope 2016，136–152）。

30. 15.  Shortly after the distribution of von Neumann’s First Draft, Eckert and Mauchly distributed a much longer—and far less famous—counter-report entitled Automatic High-Speed Computing: A Progress Report on the EDVAC (Eckert and Mauchly 1945) in which they put the emphasis on the idealized aspect the First Draft. The stakes were indeed high for Eckert and Mauchly: if the idealized depiction of the EDVAC by von Neumann was considered a realistic description of the engineering project, no patent could ever be extracted from it. And this is exactly what happened. In 1947, the Ordnance Department’s lawyers decided that the First Draft was the first publication on the project EDVAC, hence canceling the patents submitted by Eckert and Mauchly in early 1946 (Haigh, Priestley, and Rope 2016, 136–152).

31. 16.将编程视为一种应用性和常规性活动的这种观点也可以在冯·诺依曼 1946 年和 1947 年与 Arthur W. Burks 和 Herman H. Goldstine 在普林斯顿高等研究院合作撰写的更全面的报告 (Burks, Goldstine, and von Neumann 1946; Goldstine and von Neumann 1947) 中找到。在这些报告中，尤其是 1947 年的题为《电子计算仪器问题的规划和编码》的报告中，仔细考虑了科学电子计算指令序列的实现。但是，虽然要解决问题的逻辑数学规划被描述为复杂而“动态的”，但这种规划的进一步翻译主要被认为是琐碎而“静态的” (Goldstine and von Neumann 1947, 20)。编程被详细地描述为一个线性过程，在最初的规划阶段存在问题，但在实施阶段却很随意。报告没有具体说明——但这不是其目的——建模和规划阶段的错误会在实施阶段显现出来（ENIAC 投入使用时经常出现这种情况），从而使经验编程过程更加曲折而非线性。

32. 16.  This consideration of programming as an applicative and routine activity can also be found in the more comprehensive reports von Neumann coauthored in 1946 and 1947 with Arthur W. Burks and Herman H. Goldstine at Princeton Institute for Advanced Study (Burks, Goldstine, and von Neumann 1946; Goldstine and von Neumann 1947). In these reports, and especially in the 1947 report entitled Planning and Coding of Problems for an Electronic Computing Instrument, the implementation of instruction sequences for scientific electronic calculations is carefully considered. But while the logico-mathematical planning of problems to be solved is presented as complex and “dynamic,” the further translation of this planning is mainly considered trivial and “static” (Goldstine and von Neumann 1947, 20). Programming is presented, in great detail, as a linear process that is problematic during its initial planning phase but casual during its implementation phase. What the report does not specify—but this was not its purpose—is that errors in the modeling and planning phases become manifest in the implementation phase (as it was often the case when the ENIAC was put in action), making empirical programming processes more whirlwind than linear.

33. 17 . 1955 年，为了减轻 IBM 701 和即将发布的 IBM 704 的运营成本，IBM 的几位客户（其中包括 RAND 公司的 Paul Armer、洛克希德飞机公司的 Lee Amaya 和北美航空公司的 Frank Wagner）成立了一个名为“Share”的合作协会。这个客户协会以及随后成立的许多其他协会积极参与了基本程序套件的早期流通。关于这个主题，请参阅 Akera (2001; 2008, 249–274)。

34. 17.  In 1955, to alleviate the operating costs of the IBM 701 and the soon-to-be-released IBM 704, several of IBM’s customers—among them Paul Armer of the RAND Corporation, Lee Amaya of Lockheed Aircraft, and Frank Wagner of North American Aviation—launched a cooperative association they named “Share.” This customer association, and the many others that followed, greatly participated in the early circulation of basic suites of programs. On this topic, see Akera (2001; 2008, 249–274).

35. 18.有关这个名为“Whirlwind”的实时计算项目的详细历史记录，该项目最初被设计为通用飞机模拟器，请参阅 Akera (2008, 184–220)。

36. 18.  For a fine-grained historical account of this real-time computing project named “Whirlwind” that was initially designed as a universal aircraft simulator, see Akera (2008, 184–220).

37. 19.有关 SAGE 项目的更详细描述，请参阅 Redmond 和 Smith（1980、2000）、Jacobs（1986）、Edwards（1996、75–112）和 Campbell-Kelly 等人（2013、143–166）。

38. 19.  For more thorough accounts of the SAGE project, see Redmond and Smith (1980, 2000), Jacobs (1986), Edwards (1996, 75–112), and Campbell-Kelly et al. (2013, 143–166).

39. 20.根据Pugh (1995)的说法，这一合同使IBM在早期的计算机市场上占据了显著的优势。

40. 20.  According to Pugh (1995), this contract gave IBM a significant advantage on the early computer market.

41. 21 . 简而言之，瑟斯顿初级智力能力 (PMA) 测试由当时担任心理测量学会首任会长的路易斯·莱昂·瑟斯顿于 1936 年提出。该测试最初是为儿童设计的，旨在使用七个因素来衡量智力差异：词汇流畅性、言语理解能力、空间可视化能力、数字能力、联想记忆、推理能力和感知速度。有关 PMA 测试和心理测量学的简要历史，请参阅 Jones 和 Thissen (2007)。

42. 21.  In a nutshell, Thurstone Primary Mental Abilities (PMA) test was proposed in 1936 by Louis Leon Thurstone, by then the first president of the Psychometric Society. Originally intended for children, the test sought to measure intelligence differentials using seven factors: word fluency, verbal comprehension, spatial visualization, number facility, associative memory, reasoning, and perceptual speed. For a brief history of the PMA test and psychometrics, see Jones and Thissen (2007).

43. 22. EDSAC 项目的一个重要见解是使用程序的新概念来初始化系统，并使其将进一步的程序从非二进制指令转换为由 0 和 1 组成的二进制字符串。Maurice Wilkes 的博士生之一 David Wheeler 于 1949 年编写了第一个程序，他称之为“初始命令”（Richards 2005）。这种程序的功能是将其他程序转换为二进制（阴极射线管、磁芯或微处理器可以交互的唯一代码），很快就被称为“汇编程序”，并被赋予“翻译”和“语言”等语言术语（Nofre、Priestley 和 Alberts 2014）。20 世纪 50 年代，随着多家制造商投资电子计算机市场，设计了许多不同的汇编程序，从而产生了重要的兼容性问题：由于（几乎）每台新计算机对累加器和乘法器寄存器的组织方式都略有不同，因此通常需要一个新的汇编程序。问题在于汇编程序与其硬件之间的一对一关系。由于汇编程序对一个硬件操作只有一条指令，因此硬件操作组织的每次修改都需要一个新的汇编程序。然而——这是 Grace Hopper 和 IBM 的 John Backus 提出的关键见解（Campbell-Kelly 等人，2014 年，第 167-188 页）——如果我们可以提供一个更复杂的程序来将代码行转换为具有某种等效机器指令的另一个程序，而不是与硬件具有一对一关系的程序，那么我们也许能够稳定计算机编程语言，因为任何实质性的硬件修改都可以集成到位于程序员代码和硬件之间的“转换器”程序中。这是编译器的基本思想，编译器将用所谓的高级计算机语言编写的程序作为输入，并输出另一个程序——通常称为“可执行文件”，其内容可以与特定硬件交互。 20 世纪 50 年代末，FORTRAN 或 COBOL 等第一批高级计算机编程语言除了可读性更高之外，相对于汇编语言，其巨大优势还在于编译器，编译器的持续维护可以补偿和“吸收”硬件的频繁修改。例如，如果两台不同的计算机都配有 FORTRAN 编译器（这是一项关键且昂贵的条件），那么尽管两台计算机的内部组织不同，但相同的 FORTRAN 程序仍可以在这两台计算机上运行。

44. 22.  One important insight of the EDSAC project was to use the new concept of program to initialize the system and make it translate further programs from nonbinary instructions into binary strings of zeros and ones. David Wheeler, one of Maurice Wilkes’ PhD students, wrote in 1949 such very first program he called “Initial Orders” (Richards 2005). This type of program whose function was to transform other programs into binary (the only code cathode-ray tubes, magnetic core, or microprocessors can interact with) were soon called “assemblers” and cast to linguistic terms such as “translation” and “language” (Nofre, Priestley, and Alberts 2014). During the 1950s, as multiple manufacturers invested in the electronic computer market, many different assemblers were designed, thereby creating important problems of compatibility: as (almost) every new computer organized the accumulator and multiplier registers slightly differently, a new assembler was generally required. The problem lay in the one-to-one relationship between an assembler and its hardware. Since an assembler had one instruction for one hardware operation, every modification in the operational organization of the hardware required a new assembler. Yet—and this was the crucial insight of Grace Hopper and then John Backus from IBM (Campbell-Kelly et al. 2014, 167–188)—if, instead of a program with a one-to-one relationship with the hardware, one could provide a more complex program that would transform lines of code into another program with somehow equivalent machine-instructions, one may be able to stabilize computer programming languages since any substantial modification of the hardware could be integrated within the “transformer” program that lay in between the programmer’s code and the hardware. This is the fundamental idea of compilers, programs that take as input a program written in so-called high-level computer language and outputs another program—often called “executable”—whose content can interact with specific hardware. In the late 1950s, besides their greater readability, a tremendous advantage of the first high-level computer programming languages such as FORTRAN or COBOL over assembly language lay in their compilers whose constant maintenance could compensate and “absorb” the frequent modifications of the hardware. For example, if two different computers both had a FORTRAN compiler—a crucial and costly condition—the same FORTRAN program could be run on both computers despite their different internal organizations.

45. 23. 1964 年至 1967 年间，IBM 投入巨资为其计算机 System 360 开发操作系统。这个庞大的软件项目积压了大量工作，漏洞百出，管理费用昂贵，以至于其前任经理 Frederick Brooks 称之为“一个价值数百万美元的错误”（Brooks 1975）。

46. 23.  Between 1964 and 1967, IBM invested heavily in the development of an operating system for its computer System 360. The impressive backlogs, bugs, and overheads of this colossal software project made Frederick Brooks—its former manager—call it “a multi-million-dollar mistake” (Brooks 1975).

47. 24 . 1968 年，信息通用公司联合创始人 Werner Frank 发表了一篇文章，文章中提出了这样一种观点：在不久的将来，软件生产成本将超过计算机硬件成本 (Frank 1968)。尽管这种说法在许多方面都具有推测性，但直到 20 世纪 80 年代，评论家们仍在不断地重复使用和夸大其词。尽管 Frank 本人后来承认，他无意中创造了一个神话 (Frank 1983)，但这个故事“强化了人们普遍认为程序员的生产力落后的观念，尤其是与计算机硬件的惊人进步相比”(Abbate 2012, 93)。

48. 24.  In 1968, an article by cofounder of Informatics General Corporation Werner Frank popularized the idea that the cost of software production will outpace the cost of computer hardware in the near future (Frank 1968). Though speculative in many respects, this claim was fairly reused and embellished by commentators until the 1980s. Though Frank himself later acknowledged that he unintentionally generated a myth (Frank 1983), this story “reinforced a popular perception that programmer productivity was lagging, especially compared to the phenomenal advances in computer hardware” (Abbate 2012, 93).

49. 25. “逻辑语句性能”这一主题在计算机编程行为研究中反复出现，尤其是在 20 世纪 70 年代。这与 Edsger Dijkstra 引发的一场争论有关，该争论涉及高级计算机编程语言（如 BASIC 或早期版本的 FORTRAN）允许使用的 GOTO 语句（Dijkstra 1968）。根据 Dijkstra 的说法，这些在程序内部创建“跳转”的分支语句使错误定位变得非常繁琐，因此应避免。随后，他提出了“结构化编程”，这种方法包括将程序细分为更短的“模块”，以便更有效地维护（Dijkstra 1972）。20 世纪 70 年代的计算机编程行为研究通常试图评估这种方法所宣称的好处。

50. 25.  The topic of “logical statement performances” is recurrent in behavioral studies of computer programming, especially during the 1970s. This has to do with a controversy initiated by Edsger Dijkstra over the GOTO statement as allowed by high-level computer programming languages such as BASIC or early versions of FORTRAN (Dijkstra 1968). According to Dijkstra, these branch statements that create “jumps” inside a program make the localization of errors extremely tedious and should thus be avoided. He then proposed “structured programming,” a methodology that consists in subdividing programs in shorter “modules” for more efficient maintenance (Dijkstra 1972). Behavioral studies of computer programming in the 1970s typically tried to evaluate the asserted benefits of this methodology.

51. 26.为了证明第二不完备定理，哥德尔首先必须证明任何句法命题都可以用数字来表达。图灵 1937 年的证明高度依赖这一开创性的见解。关于哥德尔不完备定理与图灵关于判定性问题的命题之间的联系，请参阅 Dupuy (1994, 22–30)。

52. 26.  To prove his second incompleteness theorem, Gödel first had to show that any syntaxic proposition could be expressed as a number. Turing’s 1937 demonstration highly relied on this seminal insight. On the links between Gödel’s incompleteness theorem and Turing’s propositions regarding the Entscheidungsproblem, see Dupuy (1994, 22–30).

53. 27.神经网络，尤其是那些被定义为“深度”和“卷积”的神经网络，最近受到了广泛关注。然而，值得注意的是，神经网络的概念最初由 McCulloch 和 Pitts（他们更喜欢使用“神经元网络”的概念）在其 1943 年的论文中提出，后来被冯·诺依曼在其 1945 年的报告中采用，这与目前接受的神经网络概念大不相同。正如 Cardon、Cointet 和 Mazières ( 2018) 所表明的那样，McCulloch 和 Pitts 的神经网络最初是逻辑激活函数，后来由 Donald O. Hebb (1949) 研究，他将其与学习的概念联系起来，而学习的概念本身又被 Frank Rosenblatt (1958, 1962) 和他的感知器概念等人重新研究。马文·明斯基 (Marvin Minsky) 提出的推理规则的渐进概率化 (Minsky and Papert 1970)、反向传播算法 (Werbos 1974；LeCun 1985；Rumelhart、Hinton 和 Williams 1986) 和玻尔兹曼机 (Hinton、Sejnowski 和 Ackley 1984) 方面的研究，都积极参与了“卷积” (LeCun 等人 1989) 概念的联想，以及最近的“深度” (Krizhevsky、Sutskever 和 Hinton 2012) 概念的联想。“神经网络”一词可能在这一翻译过程中幸存下来，但它现在指的是完全不同的世界制定程序。在第 6 章的末尾，我将讨论与机器学习和人工智能相关的这一主题。

54. 27.  Neural networks, particularly those defined as “deep” and “convolutional,” have recently been the focus of much attention. However, it is important to note that the notion of neural networks as initially proposed by McCulloch and Pitts (who preferred to use the notion of “networks of neurons”) in their 1943 paper, and later taken up by von Neumann in his 1945 report, is very different from its current acceptance. As Cardon, Cointet, and Mazières (2018) have shown, McCulloch and Pitts’s neural networks that were initially logical activation functions were worked on by Donald O. Hebb (1949) who associated them with the idea of learning, which was itself reworked by, among others, Frank Rosenblatt (1958, 1962) and his notion of Perceptron. The progressive probabilization of the inference rules suggested by Marvin Minsky (Minsky and Papert 1970), the works on the back-propagation algorithm (Werbos 1974; LeCun 1985; Rumelhart, Hinton, and Williams 1986) and on Boltzmann machines (Hinton, Sejnowski, and Ackley 1984) then actively participated in the association of the notions of “convolution” (LeCun et al. 1989) and, more recently, “depth” (Krizhevsky, Sutskever, and Hinton 2012). The term “neural network” may have survived this translation process but it now refers to very different world-enacting procedures. At the end of chapter 6, I will consider this topic related to machine learning and artificial intelligence.

55. 28 . “扩展事物”和“思维事物”之间的区分在很大程度上源自笛卡尔二元论。有关笛卡尔难题的深入讨论，请参阅达马西奥 (2005) 的著作。

56. 28.  The division between “extended things” and “thinking things” derives, to a large extent, from Cartesian dualism. For thorough discussions of Descartes’s aporia, see the work of Damasio (2005).

57. 29。正如我们在第 2 章中看到的，图像处理中的显著性检测直接面临着这个问题。因此需要用适当的基本事实仔细地构建和限制显著性问题。

58. 29.  As we saw in chapter 2, saliency detection in image processing is directly confronted with this issue. Hence the need to carefully frame and constrict the saliency problem with appropriate ground truths.

59. 30.这些批评者可以追溯到古希腊智者派（Cassin 2014）。詹姆斯（1909）和梅洛-庞蒂（2013）也是重要的反对者。在发展心理学中，维果茨基（1978）提出的“社会发展理论”也是对认知主义的激烈批评。

60. 30.  One may trace these critics back to the Greek Sophists (Cassin 2014). James (1909) and Merleau-Ponty (2013) are also important opposition figures. In developmental psychology, the “social development theory” proposed by Vygotsky (1978) is also a fierce critic of cognitivism.

实证材料的呈现

Presentation of the Empirical Materials

开发用于学术出版的图像处理算法是一个涉及许多不同活动和情况的过程。但是，除了收集相关数据、构建基本事实、制定输入数据和输出目标之间的转换关系以及帮助所有这些事情发生的无数小组会议、非正式讨论、研讨会和咖啡休息之外，还有一些或多或少漫长的计算机编程过程，在这些过程中，必须编写编号的指令列表，以使电子设备能够充分计算数字数据。正是这些有始有终的行动过程，我将在本案例研究中尝试解释它们。

The development of an image-processing algorithm intended for academic publication is a process that involves many different activities and situations. But along the gathering of relevant data; the construction of ground truths; the formulation of transformative relationships between input-data and output-targets; and the numerous Group meetings, informal discussions, seminars, and coffee breaks that help all these things to happen, there are more or less long computer programming episodes when numbered lists of instructions have to be written in order to make an electronic device adequately compute digital data. It is these courses of action that have a beginning and an end that I will try to account for in this case study.

在实验室进行民族志研究的过程中，我很快发现一个问题，那就是如何记录这些行动过程。首先，由于编程过程中编写的代码非常神秘，一开始很难了解发生了什么。其次，屏幕上这些神秘符号的配置不断变化；新字符被添加，其他字符被删除，其他字符被更正，等等。第三，这些情况对参与其中的人来说似乎相当有吸引力，这让我无法询问他们在做什么。在这些看起来特别紧张的时刻，我显然格格不入。

The problem that quickly stood out during my ethnographic endeavor within the Lab was how to document these courses of action. First, as the code being written during programming episodes was very cryptic, it was in the beginning difficult to have a grip on what was going on. Second, the configurations of these cryptic signs on the screens were constantly changing; new characters were added, other erased, other corrected, and so on. Third, these situations appeared quite engaging for the people involved, which prevented me from asking them questions about what they were doing. During these moments that looked particularly intense, I was clearly out of place.

为了缓解这些方法论问题，我在实验室成员的帮助下设计了自己的图像处理项目。经过几次实验室会议后，我们一致决定，我应该尝试设计一个预处理模型，该模型可以对像素配置更合适的图像进行排序实验室正在开发一些特定的分割过程。这个不起眼的项目明确地是为了迫使我学习几种计算机编程语言的基础知识，并更熟悉图像处理。重要的是，该项目还包括一个“帮助条款”，当我陷入编程困境时，它允许我向实验室成员寻求帮助。这种有点不寻常的方法结果非常有成效。它首先让我对几种编程语言更加熟悉；¹渐渐地，所有这些神秘的符号开始变得更有意义。它也让实验室成员在我试图记录和说明的编程事件中更加自在。由于该项目是集体设计的，可能会用于未来的项目，实验室成员发现它有些相关。而且由于所谓的帮助课程与他们自己的项目没有直接关系，他们也对我在编程时做笔记和提问感到更自在。最后，也许更重要的是，这种方法让我能够更好地装备和记录编程过程：除了记录在我旁边编程的人的动作和手势外，我还可以录制我的显示器的视频并录制讨论的音频。对于这个项目所需的八次帮助会议，我最终得到了可以彻底分析的描述、屏幕录像和音频记录。

To palliate these methodological issues, I designed my own image-processing project with the help of the Lab’s members. After several Lab meetings, we collectively decided that I should try to design a preprocessing model that could sort images whose pixel configurations would fit further specific segmentation processes that were under development within the Lab. This modest project was explicitly designed to force me learn the basics of several computer programming languages and become more familiar with image processing in general. Importantly, the project also included a “helping clause” that allowed me to ask members of the Lab for help when I was stuck in a programming impasse. This somewhat unusual method turned out immensely fruitful. It first made me become more comfortable with several programming languages;¹ little by little, all these cryptic signs started to make more sense. It also made the members of the Lab more comfortable during the programming episodes I tried to document and account for. As the project had been designed collectively and could potentially be used for future projects, the members of the Lab found it somewhat relevant. And as the so-called helping sessions did not directly concern their own projects, they also felt more at ease with me taking notes and asking questions while they were programming. Finally—and perhaps more importantly—this method allowed me to better equip and document programming episodes: along with notes describing the movements and gestures of the one who was programming next to me, I could video record my monitors and audio record the discussions. For the eight helping sessions I needed for this project, I then ended up with descriptions, screen recordings, and audio recordings I could thoroughly analyze.

虽然这些帮助课程收集的材料在许多方面都很有见地，但它们仍然有局限性。由于这些课程产生的小程序主要供我自己使用，因此它们并非直接设计用于在专业程序员社区中传播，而这在企业软件环境中通常是这样的。从这个意义上说，像 Button 和 Sharrock (1995) 在其关于计算机编程实践的论文中考虑的程序阅读以就地塑造可理解性等重要主题无法得到专门研究。不过，正如我们将在本章后面看到的那样，我的一些分析命题很可能与 Button 和 Sharrock 的结论有关。

Though insightful in many respects, the materials collected during these helping sessions nonetheless had limitations. As the small programs resulting from these sessions were primarily intended for my own specific use, they were not directly designed to circulate within a professional community of programmers as it is typically the case in corporate software settings. In this sense, important topics such as program reading for the in situ shaping of intelligibility, as considered by Button and Sharrock (1995) in their paper on computer programming practices, could not be specifically investigated. Nevertheless, as we will see later in the chapter, some of my analytical propositions may well be related to Button and Sharrock’s conclusions.

以下材料取自一次帮助会议，在此期间，实验室的博士生 DF 编写了一个小程序，我将从现在起称之为 PROG，该程序处理我之前通过众包任务收集的数据。众包任务分为十轮。每轮，向 20 到 30 名未知工人展示 50 张“自然图片”，包括风景、人脸、鸟类、建筑物等。内容这些图片的种类极其多样。对于每张图像，每位工人都被要求在最先吸引他们注意力的部分周围画一个或多个矩形。在切换到下一张图像之前，每位工人还必须对他们选择标记图像特定部分的简单程度进行一到七的评分。经过十轮这项众包任务之后，254 位不同的工人每人标记了五十张图像，总共五百张图像。通过 Web 应用程序从工人的活动中收集的数据（他们处理的图像的 ID、他们画的矩形的坐标以及他们为每个标记任务给出的分数）被收集在 .txt 文件中，其组织方式如图4.1所示。这些 .txt 文件的内容以及用于众包任务的自然图像就是 PROG 必须处理的数据。

The following materials are taken from one helping session during which DF—a PhD student of the Lab—wrote a small program that I will from now on call PROG that dealt with data I had previously collected via a crowdsourcing task. The crowdsourcing task was divided into ten rounds. For each round, twenty to thirty unknown workers were shown fifty “natural pictures” of landscapes, faces, birds, buildings, and so on. The content of these pictures was extremely varied. For each image, each worker was asked to draw one or several rectangles around the parts of the image that first attracted their attention. Before switching to the next image, each worker also had to grade from one to seven how straightforward it had been for them to choose what specific parts of the image to label. After the ten rounds of this crowdsourcing task, 254 different workers each labeled fifty images for a total of five hundred images. The data collected from the activity of the workers (the IDs of the images they processed, the coordinates of the rectangles they drew, and the grades they gave for each labeling task) via a web application were gathered in .txt files organized as in figure 4.1. The content of these .txt files along with the natural images used for the crowdsourcing task were the data on which PROG had to work.

如果这个小项目明确设计用于更好地记录编程实践，那么它还有一个图像处理目标。这个次要目标是找到自然图像内容（就数字像素值的排列而言）与工人提供的矩形和等级之间的对应关系。简而言之，假设对于等级高且矩形非常分散的图像，将其内容分成更小的部分可能没有意义。对称地，对于等级低且矩形非常紧凑的图像，将其内容分成更小的部分最终可能有意义（见图4.2）。能够自动对内容可以或不可以分成更小部分的图片进行排序可能对基于分割过程的进一步有损压缩方案有用。从这个意义上说，我试图定义的计算方法最终可以作为实验室成员当时正在开发的进一步更复杂的分割/压缩方法的预处理步骤。但无论如何，要提出这样一种预处理方法，必须组装许多中间程序（包括 PROG）。

If this small project was explicitly designed to better document programming practices, it also had an image-processing goal. This secondary goal was to find correspondences between the contents of the natural images—in terms of arrangement of numerical pixel-values—and both the rectangles and grades provided by the workers. In short, the assumption was that for images with high grades and very dispersed rectangles, it may not make sense to divide their content into smaller parts. Symmetrically, for images with low grades and very compact rectangles, it may eventually make sense to divide their content into smaller parts (see figure 4.2). Being able to automatically sort pictures whose contents may or may not be divided into smaller parts could be useful for further lossy compression schema based on segmentation processes. In that sense, the computational method I tried to define could eventually serve as a preprocessing step for further, more complex, segmentation/compression methods that members of the Lab were developing at that time. But at any rate, to propose such a preprocessing method, many intermediary programs—including PROG—had to be assembled.

图 4.2中所示的支持众包任务和数据收集的 Web 应用程序的设计需要完成许多不同的程序。首先，必须设计一个 Python Web 抓取程序，以便浏览和下载 Flickr 网站 API 提供的异构、高清和 Creative Commons 许可的图像。这个小而又不简单的程序的设计首先需要与实验室成员进行“帮助会议”。其次，使用 html、JavaScript 和 PHP 计算机编程语言的几个程序必须设计为允许工人与特定数量的图像进行交互，并将其 ID、标签和成绩存储在 .txt 文件中。这个 Web 应用程序的设计需要实验室成员的两次“帮助会议”。第三，需要第一个 Matlab 程序来读取所有 .txt 文件的文本和数字内容，并在 Matlab 软件环境中重新组织它们。由于 Matlab 能够灵活地设计线性代数问题——所有整数都被视为标量——因此被广泛用于计算机科学、电子工程和经济学的研究和工业用途。然而，如果说 Matlab 编程语言以适合矩阵和数组的计算而闻名，那么它也因不适合将 .txt 数据重组为矩阵和数组而闻名。实验室成员通常将这种将数据重组为矩阵和数组的过程称为“解析”。再次，需要第四次帮助会话来帮助我组装解析程序，以进一步实现如图4.2所示的视图。

The design of the web application that enabled the crowdsourcing task and the gathering of data as shown in figure 4.2 required the completion of many different programs. First, a Python web-scrapping program had to be designed in order to browse and download heterogeneous, high-definition, and Creative-Commons-licenced images made available by the API of the Flickr website. The design of this small yet not-so-trivial program first required a “helping session” with a member of the Lab. Second, several programs using html, JavaScript, and PHP computer programming languages had to be designed to allow workers to interact with a specific number of images and store their IDs, labels, and grades within .txt files. The design of this web application required two “helping sessions” with members of the Lab. Third, a first Matlab program was required in order to read the textual and numerical contents of all the .txt files and reorganize them within Matlab software environment. Because of its agility to design problems of linear algebra—all integers being considered scalars—Matlab is widely used for research and industrial purposes in computer science, electrical engineering, and economics. Yet if Matlab programming language is known for being well adapted for the computation of matrices and arrays, it is also known for being badly adapted for the reorganization of .txt data into matrices and arrays. This reorganization of data into matrices and arrays was generally called “parsing” by the members of the Lab. Again, a fourth helping session was required to help me assemble parsing programs that further enabled views such as those presented in figure 4.2.

我们即将介绍的程序 PROG 处理由以前的解析程序重新组织的数据的分析。PROG 的形成需要 DF 的第五次“帮助”。PROG 的规格可以总结如下：出于我们将在下一节中详细介绍的原因，PROG 应该能够将图 4.2中所示的每个带标签的数字图像转换为图 4.3中所示的另一个不太复杂的数字图像。每个像素的值较不复杂的图像应对应于每个像素所属的矩形数量。例如，如果给定像素是零矩形的一部分，PROG 应将值零归因于此像素。但如果另一个给定像素是六个矩形的一部分，PROG 应将值六归因于此像素。因此，PROG 旨在将不同的值（自然图像的尺寸、众包任务参与者绘制的每个矩形的尺寸、每个像素的递增值）聚集在一起，以创建新图像，或者通常在图像处理中称为新矩阵。

The program whose formation we are about to follow—PROG—dealt with the analysis of the data as reorganized by previous parsing programs. The shaping of PROG required a fifth “helping session” with DF. The specifications of PROG can be summarized as such: for reasons we will cover at length in the next sections, PROG should be able to transform each labeled digital image as presented in figure 4.2 into another less complex digital image as presented in figure 4.3. The value of the pixels of each less complex image should correspond to the number of rectangles each pixel is part of. For example, if a given pixel is part of zero rectangle, PROG should attribute the value zero to this pixel. But if another given pixel is part of, say, six rectangles, PROG should attribute the value six to this pixel. PROG was thus intended to gather together different values (dimensions of the natural image, dimensions of each rectangle drawn by the participants of the crowdsourcing task, incrementing values of each pixel) in order to create new images or, as usually coined in image processing, new matrices.

此时，无需完全理解 PROG 的目标和规范，因为我们将在下一节中仔细考虑它们。现在更重要的是要了解 PROG 是在 Matlab 软件环境中设计的。与其他流行的高级编程语言（例如 Python 或 C）一样，Matlab 通常与包含可视化和文件组织功能的集成开发环境(IDE) 结合使用（见图4.4）。但与 Python、C 及其一些兼容 IDE（例如 PyCharm、Eclipse）不同，Matlab（作为一种独立的编程语言和IDE）由 MathWorks Inc. 拥有和维护，并以许可证为基础分发。在进行此调查时，Matlab 的专有功能受到越来越多实验室成员的批评，他们倾向于选择 Python，因为它是开源的，并得到活跃的开发者社区的支持。然而，尤其是由于其内部组织是为矩阵处理而设计的，Matlab 过去和现在都经常被使用。为了便于阅读，我对 PROG 实际形成的后续工作将仅关注 Matlab IDE 的编辑器和命令窗口。在接下来的部分中，图 4.4的内容将以图 4.5 的形式呈现。

At this point, it is not necessary to fully understand the goals and specifications of PROG as we will closely consider them in the next sections. What is more important for now is to understand that PROG was designed in the Matlab software environment. Like other popular high-level programming languages, such as Python or C, Matlab is generally used in conjunction with an integrated development environment (IDE) that includes visualization and file organization functionalities (see figure 4.4). But unlike Python, C, and some of their compatible IDEs (e.g., PyCharm, Eclipse), Matlab—as a programming language in its own right and as an IDE—is owned and maintained by MathWorks Inc. and is distributed on a license basis. At the time of this inquiry, Matlab’s proprietary feature was criticized by a growing number of Lab members who tended to prefer Python, which is open-source and supported by an active community of developers. However, notably because of its internal organization natively designed for matrix processing, Matlab was and still is frequently used. For reasons of readability, my follow-up of the practical formation of PROG will only focus on the Editor and the Command Window of the Matlab IDE. In the next sections, the content of figure 4.4 will then be presented as in figure 4.5.

即使 PROG 是该项目中最小的程序，我也无法解释它的整个形成过程。我不会解释建立 PROG 的整个编程过程，而是只关注特别有启发性的特定序列。我对编程序列的跟踪是按时间顺序进行的，从时间 0 (T0) 开始，到时间n结束。然而，每个 T 的采样并不遵循固定的时间段，而是遵循编辑器和命令窗口的修改。例如，让我们假设图 4.5是我们正在遵循的编程序列 (T0) 中 PROG 的第一个表达式。一旦程序员在编辑器和命令窗口中做出更改，这些更改将被记录并突出显示，如图4.6所示。

Even if PROG was by far the smallest program of the project, I will not be able to account for its entire formation process. Instead of accounting for the whole programming episode that established PROG, I will only focus on specific sequences that are particularly instructive. My follow-up of the programming sequences is chronological, starting at Time 0 (T0) and ending at Time n. Yet the sampling of each T does not follow a fixed period of time but rather the modifications of both the Editor and the Command Window. Let us assume, for example, that figure 4.5 is the first expression of PROG during the programming sequence we are following (T0). As soon as the programmer makes changes in both the Editor and the Command Window, these changes will be documented and highlighted as in figure 4.6.

在不同的 T 之间，程序员 (DF) 和我 (FJ) 的言论和行为将被转录。为了让内容更易读，我可能会省略一些小动作，例如快速输入错误或犹豫不决。在 T1（图 4.6）之后，编程序列将例如这样进行：

In between the different Ts, the sayings and actions of the programmer (DF) and me (FJ) will be transcribed. To keep things readable, I may omit some small actions, such as quick mistypes or hesitation disfluencies. Following T1 (figure 4.6), the programming sequence would, for example, go on like this:

国防军：  “哼，不行了。”

DF:  “Hum, it doesn’t work anymore.”

缩略词：  “显然。… ”

FJ:  “Apparently. …”

国防军：  “嘶嘶嘶嘶。”

DF:  “Tssssss.”

[在第 14 行，DF 删除“{1}”]

[at line 14, DF deletes “{1}”]

[DF运行脚本]

[DF runs the script]

[图 4.7 —T2]

[figure 4.7—T2]

国防军：  “好的。但是为什么只有两个？我不明白。今天很难！”

DF:  “OK. But why are there only two of them? I don’t get it. Difficult today!”

[笑]

[laughs]

时不时地，我也会介入以澄清事实并分析正在发生的事情。在开始第一个序列之前，重要的是要记住，我们不需要理解转录中所说的所有内容，也不需要理解每个 T 中的所有元素。在对计算机编程实践进行这种仔细分析时，重要的是每个 T 之间发生了什么。通过关注每个 T 之间的相对差异，我们将设法理解这些非常规行动过程中所涉及的一些问题。

Here and then, I will also intervene to clarify things and analyze what is happening. Before we start with the first sequence, it is important to keep in mind that one does not need to understand everything that is said in the transcriptions nor all the elements within each T. What is important in this close analysis of computer programming practices is what is happening in between each T. It is by focusing on the relative differences between each T that we will manage to understand some of the issues at stake during these unconventional courses of actions.

在我们深入探讨计算机编程实践之前，我需要提最后一件事。人们可能很容易反对，认为以下案例研究及其随后的尝试性命题并不代表一般的编程实践。对此，我的回答是，代表性在这里根本不重要。代表性确实是一个强大而重要的概念，但只有在明确界定人口边界的情况下才如此。城镇的居民、组织的细胞、书中的文字：所有这些都可以与一个非常昂贵且装备精良的集合相关——城镇的行政和地理界限、样本的物理界限、书籍的精装本——随后定义了领土和人口。在这些特定但非常罕见且经常引起争议的案例中，代表性的概念可用于提取具有统计意义的结果。但是，当没有领土、没有集合时，代表性的概念本身就失去了存在的理由。什么是编程？程序员在编程时是谁？我们不知道，因为对计算机编程实践的研究很少。这通常是民族志可以发挥作用的地方：探索未定义的——或定义有问题的——领土可能为设计后续要进行统计探索的边界提供依据。虽然我确实认为，正在编写小型 JavaScript 程序来为她个人网站菜单制作动画的莱比锡年轻街头艺术家、正在为客舱增压模块进行 Ada 最新更新的波音工程师或试图使用 Matlab IDE 解析 .txt 文件的计算机科学家在很多方面都不同——他们有不同的问题、影响、环境、设备——但我还认为 (几乎)这些情况都尚未得到民族志学的解释。我们仍然需要从某个地方开始。因此，我希望以下案例研究是更系统地研究编程行动方案的第一步；因此它的命题具有探索性。

I need to mention one last thing before we dive into the practices of computer programming. One may easily object that the following case study and its subsequent tentative propositions are not representative of programming practices in general. To this, I answer that representativeness is simply not at stake here. Representativeness is indeed a powerful and important concept but only when the boundaries of a population are clearly defined. Inhabitants of a town, cells of a tissue, words of a book: all can be related to a very costly and equipped set—the administrative and geographical limits of a towns, the physical limits of a sample, the hardcover of a book—that subsequently defines a territory and a population. In these specific—but very rare and often controversial—cases, the concept of representativeness can be used to extract statistically meaningful results. But when there is no territory, no set, the very notion of representativeness loses its raison d’être. What is programming? Who are programmers when they program? We do not know as there were very few studies of computer programming practices. This is typically where ethnography can be useful: the exploration of nondefined—or problematically defined—territories may provide takes for the design of subsequent boundaries to be explored statistically. And while I do think that the young street artist in Leipzig who is writing a small JavaScript program to animate the menus of her personal website, the engineer of Boeing who is working on the last Ada’s update for cabin pressurization modules, or the computer scientist who tries to parse .txt files with the Matlab IDE differ in many ways—they have different problems, affects, environments, equipment—I also think that (almost) none of these situations have yet been accounted for ethnographically. We still have to start somewhere. The following case study is then one of the very first steps into, I hope, more systematic studies of programming courses of action; hence the exploratory aspect of its propositions.

对齐铭文

Aligning Inscriptions

让我们关注 PROG。基于上一节中介绍的内容，我将记录一个非常短的编程序列，该序列在实时中花费的时间不到五分钟。我将尽可能贴近格式化但经验性的材料，使用我上面介绍的演示方法以及在分析过程中在 STS 中开发的几个概念。我希望表明，一组对程序员来说非常重要的实践处理铭文的扩散和对齐，以便开辟通往远程实体的通道，同时识别 位置。希望随着本章的继续，这个奇怪的命题会变得更加清晰。现在，让我们从图 4.8和4.9开始：

Let us focus on PROG. Building on what I presented in the last section, I will document a very short programming sequence that took less than five minutes in real time. I will stay as close as possible to the formatted-yet-empirical material, using the presentation method I introduced above as well as several concepts developed in STS in the course of the analysis. My hope is to show that one set of practices that are terribly important for programmers deal with the proliferation and alignment of inscriptions in order to pave out an access to a remote entity and, simultaneously, identify a location. Hopefully, this odd proposition will become clearer as the chapter goes on. For the moment, let us start in medias res with figures 4.8 and 4.9:

[图 4.8 —T0]

[figure 4.8—T0]

[DF运行脚本]

[DF runs the script]

[图 4.9 —T1]

[figure 4.9—T1]

国防军：  “好的。所以它告诉我它不起作用。”

DF:  “OK. So it tells me it doesn’t work.”

缩略词：  “显然。”

FJ:  “Apparently.”

T0 和 T1 之间发生了什么？DF 运行脚本后，命令窗口中会出现一个红色（此处为灰色）铭文，表示“索引超出矩阵维度”这段文字从何而来？谁写的？为了更好地理解这个神秘通知的来源，我必须介绍这个序列中的一个重要参与者：解释器（INT）。为了使编辑器中的十六行代码产生电脉冲，从而进一步允许计算机硬件有效地计算 .txt 文件的数据，必须采取许多步骤。幸运的是，对于我们感兴趣的情况，只有第一步是重要的。第一步是将每一行代码翻译成其他东西——在本例中是编译成机器代码的子程序——反过来，这些子程序将产生电脉冲并有效地计算数据。负责这种复杂翻译的实体之一是 INT。每次 DF 运行脚本时，都会秘密触发 INT 逐字节翻译编辑器的内容。我们不需要确切知道 INT 在翻译过程中做了什么：即使对于 DF，INT 的功能本身仍然不清楚。事实上，我们只需要了解 INT 的四个特点：

What is happening between T0 and T1? After DF runs the script, a red (here, gray) inscription appears in the Command Window, indicating that “Index exceeds matrix dimensions.” Where does this text come from? Who wrote it? To better understand the origin of this cryptic notification, I have to introduce an important participant to the sequence: the interpreter (INT). For the sixteen lines of code in the Editor to generate electric pulses that would further allow the hardware of the computer to effectively compute the data of the .txt files, many steps have to be taken. Fortunately, for the case that interests us here, only the very first step is important. This first step consists in translating every line of code into something else—in this case, subroutines compiled into machine code—that would, in turn, generate electric pulses and the effective computation of the data. One of the entities responsible for this complex translation is INT. Every time DF runs the script, INT is surreptitiously triggered to translate the content of the Editor, byte by byte. We do not need to know exactly what INT does during its translating processes: even for DF, the very functioning of INT remains obscure. In fact, we just need to understand four characteristics of INT:

1. INT 有自己的发展轨迹，几乎没有人能完全理解它：MathWorks 公司雇佣的高度专业化的团队（Matlab 的编辑）负责塑造它，并且目前仍在维护它。从这个意义上说——至少从 DF 的角度来看——INT 可以被视为一种冒着生存风险的存在（James [1912] 2003；Latour 2013），就像猫或海象一样。
1. INT has its own trajectory that is fully understood by almost nobody: highly specialized teams employed by the company MathWorks, editors of Matlab, were required to shape it and are still currently maintaining it. In that sense—at least from the point of view of DF—INT can be considered a being that takes the risk of existence (James [1912] 2003; Latour 2013), just as a cat or an elephant seal.
2. INT 逐行翻译编辑器中的每行^。2
3. INT translates one line of the Editor after the other.²
4. 一旦 INT 成功翻译了一行，如果该行指示打印铭文，INT 就会在命令窗口中打印该铭文。
5. As soon as INT successfully translates a line, if this line instructs the printing of an inscription, INT prints this inscription in the Command Window.
6. 一旦 INT 无法翻译一行，它就会停止并在命令窗口中打印红色（此处为灰色）字样。
7. As soon as INT cannot translate one line, it stops and prints a red (here, gray) inscription in the Command Window.

这引出了我们在引言中已经遇到的重要概念铭文，我强调了这些持久、移动和可再现实体的世界生成能力。当然，铭文有很多不同的类型：书籍、WhatsApp 消息、购物清单，甚至纹身都可以被视为铭文，有些比其他更持久、更移动、更可再现（Gitelman 2014）。但无论如何，铭文都是或多或少可归因事件的翻译表现，因此至少在潜在意义上构成了环境在特定情况下提供的观点。这些铭文不是（没有连字符）用于心理计算的“真实事物”的表征。它们是可以掌握的事件的格式化再现，反过来，它们又配置了其他世界生成的观点。这就是为什么我需要在第三章末尾费力地介绍实施认知：我们现在知道，主体先于认知，文献和铭文可以被视为可能暗示其他行动的行动——从一个行动到另一个行动，就我们能够感知和理解的范围而言（Penny 2017）。

This leads us to the important notion of inscription that we have already encountered in the introduction where I emphasized the world-generative capabilities of these durable, mobile, and re-presentable entities. There are, of course, many different types of inscriptions: books, WhatsApp messages, shopping lists, or even tattooed bodies can be considered inscriptions, some being more durable, mobile, and re-presentable than others (Gitelman 2014). But in any case, inscriptions are translated manifestations of more or less attributable events and thus constitute, at least potentially, takes offered by the environment in specific situations. These inscriptions are not representations (without hyphen) of “real things” that feed mental computations. They are formatted re-presentations of events that may be grasped and, in turn, configure other world-generative takes. This is why I needed to tediously introduce enactive cognition at the end of chapter 3: as we are now aware that agency precedes cognition, documents and inscriptions can be considered no more but also no less than takes that may suggest other actions—from take to take, as far as we can perceive and make sense (Penny 2017).

铭文有时能被有认知能力的个体所掌握；有时则不能。在我们的例子中，铭文“索引超出矩阵维度”确实被 DF 掌握了。事实上，当 DF 运行脚本时，他期望在命令窗口中出现铭文。此外，正如我们现在所知道的那样，DF 很清楚，命令窗口中的任何红色铭文都表明 INT 无法翻译脚本的所有行，DF 知道铭文“索引超出矩阵维度”是与 INT 相关的事件的踪迹。

Inscriptions-takes are sometimes grasped by cognizing individuals; other times, they are not. In our case, the inscription “Index exceeds matrix dimensions” is indeed grasped by DF. In fact, as DF ran the script, he expected an inscription to appear in the Command Window. Moreover, as DF is well aware—just as we are now—that any red inscription in the Command Window manifests that INT could not translate all the lines of the script, DF knows that the inscription “Index exceeds matrix dimensions” is the trace of an event related to INT.

从这一点来看，我们能够更好地理解第一个铭文对 DF 的作用。在 T1 处，铭文“索引超出矩阵维度“是 DF 把握到的一个观点，它表明有一些东西——但是什么？——正在影响 INT 的轨迹：它告诉我它不起作用。

From this point, we are able to better understand what the first inscription does to DF. At T1, the inscription “Index exceeds matrix dimensions” is a take grasped by DF that manifests that something—but what?—is affecting the trajectory of INT: it tells me it doesn’t work.

让我们继续：

Let us continue:

国防军：  “没成功。我只想检查一下图像的大小。”

DF:  “It doesn’t go through. I’ll just check the size of the image.”

[DF 在编辑器中的 2 处创建一个新行；输入“尺寸(I)”]

[DF creates a new line at 2 in the Editor; types “size(I)”]

INT 有一个问题：它不通过脚本。但是 PROG 的哪一部分影响了 INT？目前很难确切知道。事实上，从现在起，了解 INT 发生了什么对于实现 PROG 是必要的。

INT has a problem: it doesn’t go through the script. But what part of PROG is affecting INT? At this point, it is difficult to know exactly. In fact, understanding what is happening to INT is, from now on, necessary to the realization of PROG.

对于 DF，开头的红色铭文表示（虽然很模糊）INT 受物体大小的影响。术语“超出“ 和 ”方面”红色铭文的“证明”存在这种与尺寸相关的问题。为了更好地掌握影响INT轨迹的尺寸相关问题，DF首先检查了图像的大小。为此，DF添加了一行小代码“尺寸(I)”并运行脚本，从而触发INT（图4.10 —T2）。

For DF, the initial red inscription indicates—though quite vaguely—that INT is affected by the size of something. The terms “exceeds” and “dimensions” of the red inscription attest for such a size-related problem. In order to have a better grip on what size-related problem is affecting the trajectory of INT, DF starts by examining the size of the image. To do this, DF adds the small line of code “size(I)” at the second line of the script and then runs it, thus triggering INT (figure 4.10—T2).

通过添加代码行“尺寸(I)”然后触发INT，DF使得命令窗口中出现新的铭文：

By adding the line of code “size(I)” at line 2 and then triggering INT, DF makes a new inscription appear in the Command Window:

答 =

ans =

第 I 列至第 2 列

Columns I through 2

1024,712

1024 712

第 3 列

Column 3

3

3

INT 在命令窗口中打印的这个新铭文不是红色的，因此可以视为代码的实际翻译。这是理所当然的：数十年的工程发展使 DF 确信这个新铭文是 INT 的无问题表达。但是，这个铭文是否表达了正确图像的维度？如果不是，整个脚本应该重新考虑。为了验证 INT 确实无法处理正确的图像，DF 使用第二个非红色铭文创建了第三个铭文，这次来自我：

This new inscription printed by INT in the Command Window is not red and can therefore be considered an actual translation of the code. This is taken for granted: decades of engineering developments allow DF to be certain that this new inscription is an unproblematic expression of INT. But still, is this inscription expressing the dimension of the right image? If not, the whole script should be reconsidered. To verify that INT is indeed failing to process the right image, DF uses the second non-red inscription to create a third one, this time emanating from me:

国防军：  “好的，那么尺寸是 1024 × 712。你觉得可以吗？”

DF:  “OK, so the size is 1024 × 712. Does that sound right to you?”

缩略词：  “是的，这张图片是正确的。”

FJ:  “Yes, it is correct for this image.”

国防军：  “好的。所以事情发生在之后。”

DF:  “Ok. So it’s happening after.”

口头陈述“是的，这幅图像是正确的”——它本身源自我之前在一次不成功的编程尝试中遇到的铭文——让 DF 认为非红色铭文足以指代 INT 无法处理的图像。非红色铭文和铭文衍生口头陈述的明确性进一步让 DF 推断“它发生在之后”。这里的“之后”至关重要。事实上，由于第二个铭文不是红色的，并且出现在命令窗口中红色铭文的上方，DF 可以得出结论，无论是什么影响了 INT 的轨迹，它都位于指令“尺寸(I)“他刚刚在第 2 行补充道。通过添加和阐明两个新的铭文——非红色铭文和由我的确认性口头陈述传达的铭文——DF 已经对 INT 有了更清晰的了解：影响其轨迹的因素在脚本的第二行之后。

The oral statement “Yes, it is correct for this image”—itself deriving from inscriptions I had previously produced and encountered during a former unsuccessful programming attempt—allows DF to consider that the non-red inscription refers adequately to the image INT is failing to process. The certitude emanating from the articulation of the non-red inscription and the inscription-derived oral statement further allows DF to infer that “it’s happening after.” The “after” is here crucial. Indeed, since the second inscription is not red and appears above the red inscription in the Command Window, DF can conclude that whatever is affecting the trajectory of INT, it lies somewhere after the instruction “size(I)” he has just added at line 2. By adding and articulating two new inscriptions—the non-red inscription and the inscription relayed by my confirmatory oral statement—DF already gets a clearer view on INT: what is affecting its trajectory lies after the second line of the script.

让我们继续：

Let us continue:

[DF检查图4.10 —T2的命令窗口]

[DF examines the Command Window of figure 4.10—T2]

国防军：  “啊，但它也表示颜色！典型的 Matlab。”

DF:  “Ah, but it indicates also the colors! Typical Matlab.”

[DF 将光标放在“第 3 列”在T2命令窗口中]

[DF puts the cursor on “Column 3” in T2 Command Window]

国防军：  “看到了吗？[对 FJ] 我们应该只取前两个值”R“否则，它会阻塞。”

DF:  “See? [to FJ] We should take only the first two values for “R.” Otherwise, it blocks.”

缩略词：  “因为现在'R' 有三个值吗？

FJ:  “Because now ‘R’ has three values?”

国防军：  “大概吧。”

DF:  “I guess so.”

[DF 删除第 2 行；在“新”第 2 行末尾输入“.1), 大小(I,2)”]

[DF deletes line 2; at the end of “new” line 2, he types “.1), size(I,2”]

通过检查 T2 处命令窗口中的非红色铭文，DF 注意到 INT 无法处理的图像的大小由三个值表示：“1024” “712，“和”3”这在哪里“3”从何而来？很难说。它可能来自 Matlab 对构成数字彩色图像的数据的系统考虑。事实上，这些特定的矩阵与宽度、高度和三层 RGB 值绑定在一起。大多数高级编程语言不考虑这个第三个值，因为它通常不表达有关实际尺寸的有用信息图像。但 Matlab 显然以其繁琐的方式表达了这一点，根据 DF 的说法，这可能是影响 INT 的问题的根源。

By pursuing his inspection of the non-red inscription in the Command Window at T2, DF notices that the size of the image INT fails to process is expressed by three values: “1024,” “712,” and “3.” Where does this “3” come from? Difficult to say. It may come from Matlab systematic consideration of the data that structure a digital color image. Indeed, these specific matrices are bound to a width, a height, and three layers of RGB values. Most high-level programming languages do not take into consideration this third value as it generally does not express useful information about the actual dimensions of an image. But Matlab—in its fussy fashion—apparently expresses it, and this may be, according to DF, the source of the problem affecting INT.

此时，DF 认为，他通过对三个铭文（红色铭文、非红色铭文和听觉陈述（本身是过去考虑过的书面铭文的翻译））的堆积和对齐所收集的有关 INT 轨迹的记录已经足够准确，可以完成剧本；根据 DF 的说法，基于他制作、收集和对齐的证据，INT 不支持第三个值“尺寸(I)”。指向第 3 行的有关 INT 的信息反过来可以允许修改脚本并平滑 INT 的轨迹。DF 还删除了“尺寸(I)” 在第 2 行，这主要是他作为 INT 探测工具使用的工具。然后，根据他对影响 INT 轨迹的问题现象的起源的洞察，他输入了“,1),大小(I,2” 来定义“R” 仅根据两个值：”1024“ 和 ”712”，这是第一张真实图像的情况。然后他运行脚本：

At this point, DF believes that the documentation he gathered about INT’s trajectory through the piling up and alignment of three inscriptions—the red inscription, the non-red inscription, and the auditory statement (itself being a translation of written inscriptions considered in the past)—is accurate enough to complete the script; according to DF, based on the evidences he produced, collected, and aligned, INT does not support the third value of “size(I).” This information about INT that points toward line 3 may, in turn, allow the modification of the script and smooth the trajectory of INT. DF also deletes “size(I)” at line 2 that mainly served for him as an instrument for the probing of INT. Then, in line with his insight about the provenance of the problematic phenomenon that affects the trajectory of INT, he types “,1),size(I,2” in the Editor in order to define “R” according to only two values: “1024” and “712,” for the case of the first image of the ground truth. He then runs the script:

[DF运行脚本]

[DF runs the script]

[图 4.11 —T3]

[figure 4.11—T3]

国防军：  “啊不。显然它不在这里。”

DF:  “Ah no. It’s not here, apparently.”

不幸的是，对于 DF 来说，这些修改不会改变 INT 的状态。正如我们在 T3 处的命令窗口中看到的那样（图 4.11），DF 的新INT 的触发不会导致红色铭文消失：某些东西仍在影响 INT，而且它不是由三个值而不是两个值定义的图像大小。3^用科学的表达方式来说，我们可以说“INT 受到第三个值的影响”尺寸(I)” 是一件人工制品：它不参与影响 INT 轨迹的现象。反过来，有问题的位置不是第 2 行；而是在其他地方。因此需要进行更多实验；必须制作、比较和对齐更多铭文。

Unfortunately for DF, these modifications do not change the state of INT. As we can see in the Command Window at T3 (figure 4.11), DF’s new triggering of INT does not lead to the disappearance of the red inscription: something is still affecting INT, and it was not the image size defined by three values instead of only two.³ Using a scientific expression, we can say that “INT-being-affected-by-the-third-value-of-size(I)” was an artifact: it does not participate in the phenomenon that affects INT’s trajectory. In turn, the problematic location is not line 2; it is somewhere else. More experiments are therefore needed; more inscriptions have to be produced, compared, and aligned.

神器“INT 受到第三个值的影响——尺寸(I)不过，对于 DF 来说，这并非完全没有用。多亏了它，DF 现在可以确定 INT 受到了第 2 行之后发生的与大小相关的问题的影响。但目前对 INT 的这种确定性还太薄弱；它不允许 DF 准确识别影响 INT 的因素，因此无法相应地修改代码。

The artifact “INT-being-affected-by-the-third-value-of-size(I)” was not totally worthless for DF, though. Thanks to it, DF is now certain that INT is being affected by a size-related problem that occurs after line 2. But this certainty about INT is for the moment too thin; it does not allow DF to precisely identify what is affecting INT and therefore modify the code accordingly.

让我们继续：

Let us continue:

国防军：  “好的。那么我们就打印矩形。然后比较一下。”

DF:  “OK. Well, we’ll print the rectangle then. And just compare.”

[DF 删除“；” 在第 8 行末尾；他在编辑器中的第 3 行创建一个新行；他输入“尺码(R)”第 3 行]

[DF deletes “;” at the end of line 8; he creates a new line at 3 in the Editor; he types “size(R)” at line 3]

PROG 处理的是自然图像，在众包任务期间，工人先前已在其上绘制了矩形。正如我们在上一节中介绍本章的经验材料时所看到的，绘制的矩形严格来说并不在图像上：它们作为坐标存储在 .txt 文件中。我们现在正在研究的脚本旨在使用每个自然图像及其矩形的宽度和高度值来创建一个不太复杂且更易于分析的新图像。这些新的简化图像（从现在开始我将称之为矩阵）应该只表达工人在初始彩色图像上绘制的矩形的数量和位置。在这方面，脚本的工作流程非常简单：首先，使用初始自然图像的宽度和高度值创建一个空矩阵，然后使用与该图像相关的 .txt 文件中的工人数据创建一个矩形，然后将矩形添加到空矩阵中。随着越来越多的矩形添加到矩阵中，矩阵逐渐获得更多值。在图像处理领域，我们称矩阵为递增。图 4.3提供了 PROG 最终输出的两个示例；即根据 .txt 文件中与 ID 相关的矩形的坐标递增的矩阵。但我们还没有到达那里；在编程这一集的这个阶段，INT——这个生动的至少对于双足哺乳动物来说，难以掌握的实体受到某种因素的影响，导致其无法充分翻译代码。

PROG deals with natural images on which rectangles have been previously drawn by workers during a crowdsourcing task. As we saw in the previous section that presented the empirical materials of this chapter, the drawn rectangles are not strictly speaking on the images: they are stored as coordinates within .txt files. The script we are now examining is intended to use the width and height values of each natural image as well as its rectangles in order to create a new image that is less complex and easier to analyse. These new simplified images—that I will from now on call matrices—should only express the number and the position of the rectangles that the workers drew on the initial color images. In this respect, the workflow of the script is quite straightforward: first, an empty matrix is created using the width and height values of the initial natural image, then a rectangle is created using the workers’ data in the .txt file related to this image, then the rectangle is added to the empty matrix. Progressively, as more and more rectangles are added to the matrix, the matrix acquires more values. In the field of image processing, we say that the matrix is incremented. Figure 4.3 provides two examples of PROG’s final outputs; that is, matrices that have been incremented according to the coordinates of the rectangles related to their IDs in .txt files. But we are not there yet; at this point of the programming episode, INT—this lively entity on which it is difficult to have a grip, at least for biped mammals—is affected by something that prevents it from translating the code adequately.

什么影响了 INT 尚不清楚。但 DF 之前处理和对齐的铭文让他意识到 INT 的问题与某种大小和维度有关。此外，DF 还了解脚本的一般工作流程，因为他主要设计了它（稍后会详细介绍）。在这方面，如果添加到第一个矩阵的第一个矩形超出了矩阵的边界怎么办？这将是非常成问题的，因为这意味着一些 .txt 数据已损坏。但由于矩形已索引到 .txt 数据，这将满足红色铭文“指标超出矩阵维数“但是DF怎么能确定呢？就像以前一样，通过制作更多铭文并进行比较。

What is affecting INT is not clear. But the previous inscriptions DF managed to handle and align have made him see that INT’s problem has to do with some size and dimension. Moreover, DF is also aware of the general workflow of the script since he mostly designed it (more on this later). In this respect, what if the first rectangle that is added to the first matrix exceeds the boundaries of the matrix? It would be very problematic as it would signify that some .txt data are corrupted. But as the rectangle is indexed to .txt data, this would satisfy the red inscription “Index exceeds matrix dimension.” But how could DF be certain of that? Just as before, by producing more inscriptions and compare them.

要打印第一个矩形的尺寸，DF 删除“；”在第 8 行末尾^。4为了打印数据集中第一幅图像的尺寸，他写道“尺码(R)”第 3 行。然后他运行脚本：

To print the size of the first rectangle, DF deletes “;” at the end of line 8.⁴ In order to print the dimension of the first image of the dataset, he writes “size(R)” on line 3. He then runs the script:

[DF运行脚本]

[DF runs the script]

[图 4.12 —T4]

[figure 4.12—T4]

[DF检查图4.12 —T4的命令窗口]

[DF examines the Command Window of figure 4.12—T4]

国防军：  “所以，197 和 323 相加，显然小于 1024。高度也是一样。好吧。这很奇怪，因为它没有超过。 ”

DF:  “So, 197 and 323. Makes less than 1024, obviously. And same for height. Alright. It’s strange because it doesn’t exceed.”

在 T4 处的命令窗口中出现了两个新的非红色铭文，因此先验上没有问题（图 4.12）。第一个是“答 = 1024,712”描述集合中第一幅图像的尺寸。第二幅“直角 = 197 91 323 371” 描述了第一个工人绘制的第一个矩形的尺寸以及该矩形在第一幅图像中的位置。rect 的第一个值“197”表示图像中的水平坐标，第二个值“91” 表示其垂直坐标。因此，这两个数字表示矩形从图像的 [197:91] 像素开始。rect 的第三个值“323”表示矩形的宽度，第四个值“371”表示其高度。因此，最后两个数字表示矩形的宽度为 323 像素，高度为 371 像素。

Two new non-red, and thus a priori nonproblematic, inscriptions appear in the Command Window at T4 (figure 4.12). The first one “ans = 1024 712” describes the dimension of the first image of the collection. The second one “rect = 197 91 323 371” describes the dimensions of the first rectangle drawn by the first worker as well as the location of this rectangle within the first image. The first value of rect, “197,” refers to its horizontal coordinate within the image, and the second value, “91,” refers to its vertical coordinate. These two numbers therefore indicate that the rectangle starts at pixel [197:91] of the image. The third value of rect, “323,” expresses the width of the rectangle and the fourth value, “371,” expresses its height. These two last numbers therefore indicate that the width of the rectangle is 323 pixels and that its height is 371 pixels.

在 T4 时，DF 已经知道所有这些值所指的含义；在此编程事件之前，我向他解释了我用来构造 .txt 文件数据的约定。但是，一旦打印出这些值并将其与图像的宽度和高度进行比较，就可以进行基本但非常重要的算术评估：“197 + 323 < 1024”和“91 + 371 < 712”。这些是至关重要的线索，因为它们无法证实命令窗口的红色铭文；矩形不超过图像的尺寸。矩形的大小和位置不是影响 INT 的因素。其他因素正在破坏 INT 与 PROG 的关系。但是什么？它在哪里？需要更多的铭文来更好地记录影响 INT 的因素并相应地修改脚本。

At T4, DF is already aware of what all these values refer to; before this programming episode, I explained to him the conventions I used to structure the data of the .txt files. But once these values are printed and compared with the width and height of the image, basic yet terribly important arithmetic evaluations can be undertaken: “197 + 323 < 1024” and “91 + 371 < 712.” These are crucial clues as they do not corroborate the red inscription of the Command Window; the rectangle doesn’t exceed the dimensions of the image. The size and position of the rectangle is not what is affecting INT. Something else is disrupting INT in its relation with PROG. But what? And where is it? More inscriptions are required to better document what affects INT and modify the script accordingly.

我们在 T4 中看到的是我在这里试图强调的过程的完美示例：通过打印图像的大小和矩形的坐标，DF 可以更好地掌握手头的过程。他可以清晰地表达这两个新铭文并将它们与之前的铭文对齐。从这个意义上说，他正在积极地开辟通往 INT 及其红色铭文的通道。尽管这种铭文的制作和对齐并不像 DF 希望的那样有效——矩形的尺寸没有超过图像的尺寸——但这为他提供了有关所研究现象的另一条线索：影响 INT 的因素在其他地方。我相信，这种掌握、制作和对齐铭文以确定问题现象的起源的做法是编程的核心。正如我们将看到的，它并不是在计算机编程序列中部署的唯一类型的实践。但在某些特殊情况下，当重要实体在其轨迹中受阻，从而阻止通过电脉冲计算数据时，铭文的处理和对齐仍然至关重要。在这些必须找到有问题的位置的情况下，实验的设计和其结果的表达似乎是必要的，以铺平一条非常具体的路径，本身提供了关于一些小的、分散的、非常快速的实体的非常具体的信息，我们可以称这些实体为“解释器”、“编译器”，甚至在微代码的情况下为“处理器”。我将在本编程序列结束时回到这个命题。但此时，重要的是要注意，DF 目前正在进行的铭文的平凡添加和对齐可能是计算机编程活动的核心。

What we see at T4 is a perfect example of the process I’m here trying to highlight: by printing the size of the image and the coordinates of the rectangle, DF acquires a better grip on the process at hand. He can articulate these two new inscriptions and align them to the previous ones. In that sense, he is enactively paving out some access to INT and its red inscription. Even though this production and alignment of inscriptions do not work as DF hoped—the dimensions of the rectangle do not exceed the dimensions of the image—this gives him another clue about the phenomenon under scrutiny: what is affecting INT lies somewhere else. This practice of grasping, producing, and aligning inscriptions in order to identify the origin of a problematic phenomenon is, I believe, central to programming. As we will see, it is not the only type of practices that are deployed during computer programming sequences. But in some specific situations, when an important entity is blocked in its trajectory, thus preventing the computation of data by means of electric pulses, the handling and aligning of inscriptions remains crucial. In these situations when a problematic location has to be found, the design of experiments and the articulation of their results appear necessary to pave a very specific path, itself providing very specific information about some small, scattered, and very swift entities we may call “interpreters,” “compilers,” or even “processors” in the case of microcode. I will come back to this proposition at the end of this programming sequence. But already at this point, it is important to note that the mundane addition and alignment of inscriptions DF is currently making might be central to the very activity of computer programming.

考虑到这些初步要素，让我们继续：

With these preliminary elements in mind, let us continue:

国防军：  “我会尝试其他方法。我们会看看矩形是否对应。”

DF:  “I’ll just try something else. We’ll see if the rectangle corresponds.”

[DF 在编辑器中的第 13 行创建一个新行；在这个新行上，他输入“imshow(I(y，x，:))”]

[DF creates a new line at 13 in the Editor; on this new line, he types “imshow(I(y,x,:))”]

DF 需要一个新的铭文：如果矩形和图像之间的关系对 INT 来说没有问题，那么一定是其他东西有问题。但是什么呢？就像编程过程中经常发生的那样，情况开始变得令人困惑。为了确保 T4 时命令窗口中显示的矩形是正确的，而不是某种尚未识别的伪像，DF 需要在叠加在第一幅图像上时看到这个第一个矩形。为此，他在编辑器中创建一个新行并输入小指令“imshow(I(y，x，:))”然后他运行脚本：

DF needs a new inscription: if the relationship between the rectangle and the image is not problematic for INT, something else must be. But what? As is often during programming episodes, the situation starts to be confusing. To be sure that the rectangle expressed in the Command Window at T4 is the right one and not some sort of not-yet-identified artifact, DF needs to see this first rectangle when superimposed over the first image. To do so, he creates a new line in the Editor and types the small instruction “imshow(I(y,x,:)).” He then runs the script:

[DF运行脚本]

[DF runs the script]

[图4.13 —T5]

[figure 4.13—T5]

图4.14​

[figure 4.14]

[DF 检查图 4.14 ]

[DF examines figure 4.14]

国防军：  “好的。所以从理论上讲，这应该是第一位工人标记的第一个矩形。”

DF:  “OK. So theoretically, this should be the first rectangle labeled by the first worker.”

DF 在 T5 触发的新铭文（图 4.14）这次有些不同。它不是文本，而是图像的一部分。更准确地说，它是第一个工人在第一幅图像上画的第一个矩形的表达。就像在 T2 和 T3 之间一样，这个新铭文让 DF 可以创建另一个铭文，这是从我身上散发出来的时间：

The new inscription triggered by DF at T5 (figure 4.14) is this time a little different. Instead of text, it is a part of an image. More precisely, it is the expression of the first rectangle the first worker drew on the first image. And just like between T2 and T3, this new inscription allows DF to create another inscription, this is time emanating from me:

国防军：  “对应吗？”

DF:  “Does it correspond?”

缩略词：  “是的，是的。”

FJ:  “Yes, yes, it does.”

国防军：  “好的。所以它肯定在其他地方阻塞了。也许它无法定义第二个矩形。”

DF:  “OK good. So it definitely blocks somewhere else. Maybe it can’t define the second rectangle.”

在研究了几天的地面实况数据后，我成为了一个值得信赖的参考：至少对于第一幅图像，我非常清楚不同矩形的位置。再次，两个铭文的表达和对齐——第一幅图像上的第一个矩形和我自己的验证（由我之前遇到的铭文提供信息）——允许 DF 继续探究涉及 INT 的问题现象。如果第一个矩形和负责定义它的代码部分不是影响 INT 的原因，那么问题应该出在其他地方。也许在第二个矩形中，更一般地说，在负责定义它的代码部分中？再次需要新的铭文：

Having worked on the data of the ground truth for a couple days, I am a trustworthy reference: at least for the first image, I know quite well the position of the different rectangles. Once again, the articulation and alignment of two inscriptions—the first rectangle over the first image and my own verification (informed by inscriptions I had previously encountered)—allow DF to pursue his inquiry into the problematic phenomenon engaging INT. If the first rectangle and the part of the code responsible for defining it are not what is affecting INT, the problem should lie somewhere else. Perhaps in the second rectangle and, more generally, the part of the code responsible for defining it? Once again, new inscriptions are required:

国防军：  “当我们定义空矩阵时可能是这样的。”

DF:  “It might be when we define the empty matrix.”

[DF 删除“imshow(I(y，x，:))”在第 13 行；在第 2 行，他选择了函数“零”，右键单击它，然后选择“选择帮助”]

[DF deletes “imshow(I(y,x,:))” on line 13; on line 2, he selects the function “zeros,” right clicks on it, and selects “help on selection”]

图4.15​

[figure 4.15]

新的铭文（图 4.15）与命令窗口中显示的铭文又略有不同。事实证明，Matlab IDE 确实提供了对“选择帮助”数据库的访问，如果连接到互联网，该数据库将显示每个选定函数的正确语法。此弹出窗口与第 2 行的可疑函数对齐，DF 可以使用鼠标光标将帮助菜单的正确语法与编辑器中的内容进行比较：

The new inscription (figure 4.15) is again a little different from those appearing in the Command Window. It turns out indeed that the Matlab IDE provides access to a “Help on Selection” database that, if connected to the internet, displays the correct syntax for each selected function. This pop-up window being aligned with the suspect function at line 2, DF can use the mouse cursor to compare the correct syntax of the help menu with what is written in the Editor:

国防军：  “不，不，我们做对了。它在别的地方。”

DF:  “No, no, we did it right. It is somewhere else.”

[DF 关闭“选择帮助”窗口]

[DF closes the “help on selection” window]

通过比较帮助菜单和脚本，DF 可以确定 INT 不会受到这行代码的影响；语法是正确的，因此 INT 能够理解它。问题出在其他地方：

The comparison between the help menu and the script allows DF to be certain that INT is not affected by this line of code; the syntax is right, so INT is able to understand it. The problem lies somewhere else:

[DF运行脚本]

[DF runs the script]

[图4.16 —T6]

[figure 4.16—T6]

国防军：  “呃，我不明白……只有空的矩阵。”

DF:  “Huh, I don’t get it … There’s only the empty matrix.”

在 T6（图 4.16）时，DF 有点迷失了。他刚刚制作的新铭文很难理解；它与之前的铭文有什么关系？零仅指空矩阵“R”从定义上讲，这个铭文不能变得太大。正如法律上所说，这个铭文“不符合条件”；这个铭文与之前的铭文之间没有关系。必须尝试其他方法：

At T6 (figure 4.16), DF is getting a little lost. The new inscription he has just produced is difficult to grasp; how does it relate to the previous ones? The zeros only refer to the empty matrix “R” that, by definition, cannot become too big. This inscription is “not eligible” as one says in law; no relationship between this inscription and the previous ones can be established. Something else has to be tried:

国防军：  “这太愚蠢了。抱歉，我有点生疏了……我还是试试别的方法吧。”

DF:  “It’s so stupid. Sorry, I’m a bit rusty … I’ll just try another way.”

[在第 15 行末尾，DF 输入“= R(y,x) + 个数（y 个数，x 个数）;”]

[at the end of line 15, DF types “= R(y,x) + ones(numel(y), numel(x));”]

国防军：  “所以基本上（对于 FJ），我会做一个包含一个的 1 × 1 矩阵，然后根据区域的大小重复它。这很愚蠢，但至少我确信它会起作用。我们会看看它是否能改变什么。”

DF:  “So basically [to FJ], I do a 1 × 1 matrix that contains one and then I repeat it according to the size of the region. It’s very stupid, but at least I’m sure it will work. We’ll see if it changes anything.”

[DF运行脚本]

[DF runs the script]

[图4.17 —T7]

[figure 4.17—T7]

国防军：  “好吧，至少它没有改变任何东西。它也没有阻塞这里。”

DF:  “Well, at least it doesn’t change anything. It doesn’t block here either.”

DF 的实验是结论性的。在 T6（图 4.16）时，他对第 15 行的指令并不完全信服。在 T7（图 4.17）时，他尝试了另一种等效的“愚蠢”方式来表达它。我们不需要对此进行深入研究代码的情感方面，因为我们将在本章后面讨论它。此时，更重要的是 DF 使用了一条他确信 INT 可以翻译的指令。这一事实的可靠性，肯定在他之前使用 Matlab 编程语言的经历中得到了巩固，使他能够进行新的实验。再一次，当与之前的铭文联系起来时，两个新的铭文“答 = 1024,712“ 和 ”直角 = 197 91 323 371” 很有启发性；因为它们与 T4 中出现的类似，DF 可以得出结论，涉及 INT 的问题现象并非源自脚本的第 15 行。它必须在其他地方，再次说明：

The experiment of DF is conclusive. At T6 (figure 4.16), he was not totally convinced by the instruction at line 15. At T7 (figure 4.17), he tries another equivalent “stupid” way to express it. We do not need to dig too far into this affective aspect of code since we are going to consider it later on in the chapter. At this point, what is more important is that DF used an instruction he was certain INT could translate. The solidity of this fact, certainly consolidated during his previous experiences with Matlab programming language, allows him to equip a new experiment. Once again, when articulated with the previous inscriptions, the two new inscriptions “ans = 1024 712” and “rect = 197 91 323 371” are instructive; as they are similar to the ones that appeared at T4, DF can conclude that the problematic phenomenon engaging INT does not derive from the line 15 of the script. It has to be somewhere else, again:

国防军：  “好吧，我会做一些非常非常愚蠢的事情，但我只是想看看它是否在这里。”

DF:  “OK, I’ll do something very, very stupid but I just want to see if it’s here.”

[DF 在 7 处创建新行；输入“1”；在 10 处创建新行；输入“2”]

[DF creates a new line at 7; types “1”; creates a new line at 10; types “2”]

[DF运行脚本]

[DF runs the script]

[图4.18 —T8]

[figure 4.18—T8]

[DF检查图4.18 —T8的命令窗口]

[DF examines the Command Window of figure 4.18—T8]

国防军：  “好的。它在这里[图 4.18第 9 行— T8]。看到了吗？[DF 将光标放在第 9 行] 它给出'1，' 然后 '直角，' 然后 '2，' 然后 '1”然后停了下来。就是这个”j+3' 在第一个矩形之后，它变得太大了。它采用第一个矩形，如果第二个矩形更大，它就无法增加。“

DF:  “OK. It’s here [at line 9 of figure 4.18—T8]. See? [DF puts the cursor on line 9] It gives ‘1,’ then ‘rect,’ then ‘2,’ then ‘1,’ then stops. It’s this ‘j+3’ that becomes too big after the first rectangle. It takes the first rectangle, and if the second rectangle is bigger, it just can’t increment.”

在 T8（图 4.18）处，这个愚蠢的行为得到了回报：新的铭文成功地识别出了 INT 遇到的问题现象的来源。第 9 行，“j+3”在第一个矩形之后变得太大，从而扰乱了 INT 的翻译工作。但 DF 是如何得出这个推论的？他是如何自信地将扰乱 INT 的责任归咎于第 9 行？如果我们像 DF 一样仔细观察 T8 的命令窗口，我们会看到它的第一组数字——“1024“ 和 ”712”——表示“R”正如编辑器脚本第 3 行指示的那样。如果我们继续检查，我们会看到后续数字“1” 表达指令“1”正如第 8 行指示的那样。然后我们看到第三组数字——“197” “91” “323，“ 和 ”371”——表示第 9 行指示的第一个矩形的大小。然后是命令窗口中的第四个数字——“2”——表达指令“2”按照第 10 行的指示。第五个数字——“1”——再次表达了这样的指示“1” 在第 8 行。这个元素至关重要，因为它表明，在这个特定时刻，INT 即将处理第二个矩形。正如命令窗口的最后一个元素所示，一旦 INT 尝试第二次平移第 9 行，它就会阻止并打印红色错误。通过依次检查命令窗口，我们可以识别出影响 INT 的因素（对于 DF 而言）：在脚本的第二轮，INT 无法平移第 9 行。这最后一个元素允许 DF 将与 INT 相关的现象的起源归因于一个特定位置。

At T8 (figure 4.18), the stupid thing pays off: the new inscription successfully identifies the source of the problematic phenomenon engaging INT. At line 9, “j+3” becomes too big after the first rectangle, thus disrupting INT in its translation efforts. But how does DF make this inference? How does he confidently attribute to line 9 the responsibility of disrupting INT? If we look attentively at the Command Window of T8, just as DF does, we see that its first series of numbers—“1024” and “712”—expresses the size of “R” as line 3 of the script in the Editor instructs it. If we continue our examination, we see that the subsequent number “1” expresses the instruction “1” as line 8 instructs it. Then we see that the third series of numbers—“197,” “91,” “323,” and “371”—expresses the size of the first rectangle as line 9 instructs it. Then the fourth number in the Command Window—“2”—expresses the instruction “2” as instructed at line 10. The fifth number—“1”—expresses, again, the instruction “1” on line 8. This element is crucial because it shows that, at this specific moment, INT is about to deal with the second rectangle. And as the last element of the Command Window indicates, as soon as INT tries to translate line 9 for the second time, it blocks and prints a red error. By sequentially examining the Command Window, what is affecting INT becomes for us—as for DF—identifiable: at the second round of the script, INT is not able to translate line 9. This last inscription allows DF to attribute the origin of the INT-related phenomenon to one specific location.

在这一点上，重要的是要记住，这最后的铭文——尽管至关重要——本身并不能构成联系INT 的红色铭文和第 9 行之间。正是所有先前铭文的添加和对齐逐渐导致了最后一个铭文的定义。整个对齐过程让 DF 能够确定影响 INT 的现象的出处：它无法翻译“j+3”第二次出现在第 9 行。

At this point, it is important to remember that this last inscription—even though crucial—did not allow by itself the constitution of a connection between INT’s red inscription and line 9. It is the addition and the alignment of all the previous inscriptions that progressively led to the definition of this last inscription. The whole aligning process allowed DF to pinpoint the provenance of the phenomenon affecting INT: it cannot translate “j+3” at line 9 for the second time.

有些读者可能已经注意到，为了解释这个小的编程序列，我使用了 STS 文献中提出的几个概念来描述一个先验上非常不同的过程：科学实验室中的实验实践。现在我需要讨论我暗中建立的实验室实践与计算机编程实践之间的联系。

As some readers may have noticed, in order to account for this small programming sequence I used several notions that have been developed in the STS literature to describe an a priori very different process: experimental practices in scientific laboratories. I now need to discuss this connection between laboratory practices and computer programming practices I have surreptitiously drawn.

在过去的五十年中，许多科学研究都强调了文本文献（Latour and Woolgar 1986）、图表（Netz 2003）、图形（Dennis 1989；Gooday 1990）和笔记（Lynch 1985；Garfinkel 1981）的核心地位，我在此根据 Latour (2013) 的观点将这些内容归纳为“铭文”。其他重要研究也表明了制作、对照和表达这些铭文所需的仪器和实验的核心地位（Hacking 1983；Knorr-Cetina and Mulkay 1983；Collins 1975；Dear 1987；Gooding、Pinch 和 Schaffer 1989）。还有一些研究进一步强调了这些铭文的操纵和传播的重要性（Latour 1987；Knorr-Cetina 1999），通过比较、对峙、对齐——简而言之，就是表达——有时最终会形成 Latour（1999a）所说的“参考链”：或多或少固化的路径，当一切就绪时，记录某个遥远实体（例如行星、病毒、粒子）的行为。这些重要的研究表明，经过认证的知识是既生产的又是客观的：由于科学实践——以及支持这些实践表达的科学机构——知识是客观^的。5

For the last fifty years, many studies of scientific work have underlined the centrality of textual documents (Latour and Woolgar 1986), diagrams (Netz 2003), graphs (Dennis 1989; Gooday 1990), and notes (Lynch 1985; Garfinkel 1981) that I gather here—following Latour (2013)—under the umbrella term “inscriptions.” Other important studies also showed the centrality of the instruments and experiments required to produce, confront, and articulate these inscriptions (Hacking 1983; Knorr-Cetina and Mulkay 1983; Collins 1975; Dear 1987; Gooding, Pinch, and Schaffer 1989). And still other studies further emphasized the importance of the manipulation and circulation of these inscriptions (Latour 1987; Knorr-Cetina 1999) that, through comparison, confrontation, alignment—in short, articulation—sometimes end up forming what Latour (1999a) calls “chains of reference”: more or less solidified paths that document, when everything is in place, the behavior of some remote entity (e.g., a planet, a virus, a particle). These important studies present certified knowledge as being produced and objective at the same time: thanks to scientific practices—and scientific institutions that support the expression of these practices—knowledge is objective.⁵

正如这个简短的编程序列似乎表明的那样，编程实践有时（并非总是）类似于构建认证知识所需的一些实践。事实上，通过实验和仪器制作铭文，以及为了制作更多铭文而对它们进行比较和对齐，这与科学实验室中观察到的情况非常吻合。一点一点地，通过对铭文的操作、比较和对齐，开辟了一些通道，可以用于描述涉及远程实体的现象。在计算机编程的情况下，这个远程实体可能有所不同：例如，它可以是 Matlab 解释器、C 编译器或 Intel 微处理器。无论如何，这些不同实体的共同特点是其构成关系非常迅速。事实上，如何才能掌握一个解释器、编译器，或者最糟糕的是，一个每秒执行数十亿次运算的处理器？一旦组装起来，这些实体就很难掌握；因此，科学的验证模式对于更好地了解影响它们的因素至关重要。此外，我认为在计算机编程过程中采用实验室实践并不是 20 世纪 50 年代末平面工艺发展后电子元件小型化的结果（L é cuyer、Brock 和 Last 2010）。根据对早期电子计算机的历史研究，这些计算机由两米高的累加器和乘法器组成，而这些累加器和乘法器本身由数百个通过电线和焊点连接的电阻器组成。计算过程中发生的每个短路、进位错误或分频器故障都必须通过繁琐的错误报告、铭文和实验来识别和定位（Haigh、Priestley 和 Rope 2014；2016，60–83）。在电子计算的早期，程序员还必须对齐铭文以找到系统受影响组件的访问权限。

As this short programming sequence seems to indicate, programming practices may sometimes—not always—resemble some of the practices required for the construction of certified knowledge. Indeed, the production of inscriptions—via experiments and instruments—and their comparison and alignment in order to produce even more inscriptions echo well with what has been observed in scientific laboratories. Little by little, through the manipulations, comparisons, and alignments of inscriptions, some access is paved out that may allow the characterization of a phenomenon engaging a remote entity. In the case of computer programming, this remote entity may vary: it can be, for example, a Matlab interpreter, a C compiler, or an Intel microprocessor. At any rate, the common characteristic of these different entities is the incredible swiftness of their constitutive relationships. Indeed, how is it possible to have a grip on an interpreter, a compiler or—worst—a processor that executes billions of operations per second? Once assembled, these entities are very difficult to grasp; hence the relevance of the scientific mode of veridiction to better understand what is affecting them. Moreover, I assume that the adoption of laboratory practices during computer programming episodes is not a result of the miniaturization of electronic components that followed the development of planar process at the end of the 1950s (Lécuyer, Brock, and Last 2010). As shown by historical studies of early electronic computers made of two-meter-high accumulators and multipliers—themselves made of hundreds of resistors connected with wires and soldered joints—every short circuit, carry errors, or divider fault that occurred during computation episodes had to be identified and located through the tedious formation of error reports, inscriptions, and experiments (Haigh, Priestley, and Rope 2014; 2016, 60–83). In these early days of electronic computing, programmers also had to align inscriptions to pave out an access to the affected component of the system.

科学实践和计算机编程实践之间的另一个相似之处是，人们通常倾向于忘记那些能够表征所研究现象的工具。在这两种情况下，当通过特定的实验室环境确定了某种现象的来源后，人们通常会将形成参考链的实践、工具和实验放在一边（Latour and Woolgar 1986, 105–155）。科学的这一特性使得科学史研究困难。由于既定事实是从最初使它们得以组装和固化的支架中提炼出来的，因此人们可能很容易从既定事实开始向后推断（Collins 1975）。因此，要从经验上掌握科学实践，至关重要的是将事实视为特定过程的结果，而不是先前事件的原因（Bloor 1981）。在较小程度上，计算机编程也是如此。当表征涉及远程实体的现象时；当脚本中的问题位置被确定后，大多数工具（小段代码、对 FJ 的疑问、“愚蠢的东西”）都会被放在一边，很快就会被遗忘。在编程阶段结束时，当脚本可以正常运行并按预期执行时，大多数这些中间对象（Vinck2011) 通常会被遗忘。因此，如果以完成的脚本或程序作为编程研究的起点，那么错过完成这些脚本或程序所必需的内容的风险就越大^。6

Another similarity between scientific practices and the practices of computer programming is a common tendency to forget about the instruments that enabled the characterization of the phenomenon under scrutiny. In both cases, when the source of a phenomenon has been identified thanks to a specific laboratory setting, the practices, instruments, and experiments that allowed the formation of the chain of reference are generally put aside (Latour and Woolgar 1986, 105–155). This characteristic of science can make its history difficult to conduct. As established facts are purified from the scaffoldings that allowed them to be assembled and solidified in the first place, great may be the temptation to start from established facts and extrapolate backward (Collins 1975). To empirically grasp the practice of science, it is therefore crucial to consider facts as consequences of specific processes rather than causes of prior events (Bloor 1981). To a lesser extent, the same is true for computer programming. When the phenomenon engaging the remote entity is characterized; when the problematic location in the script is identified, most of the instruments (small bits of code, questions to FJ, “stupid things”) are put aside and soon forgotten. At the end of the programming episode, when the script is functional and performs as desired, most of these intermediary objects (Vinck 2011) are generally left behind. Consequently, if one takes completed scripts or programs as starting points for the study of programming, the greater is the risk to miss what has been necessary to complete these scripts or programs.⁶

对于计算机编程的情况，人们可以想象我上面记录的对齐实践的不同表达。尽管我推测这些表达仍然在于形成引用链，以便访问远程实体并指向编号铭文列表中的特定位置，但它们可能不一定部署在与 DF 类似的时空地标中。如果我们考虑“程序测试”——一个重要的工业过程，包括检测和记录错误以修改代码行——这项工作可以在空间和时间上高度分布（Parrington 和 Roper 1989；Myers、Sandler 和 Badgett 2011）。7^当我们的某个软件程序因神秘原因崩溃时，我们经常遇到的“错误报告”是对齐铭文必要性的另一种表达，因为它们恰恰在于记录程序在什么时间以及在什么操作之后对解释器、编译器或处理器产生了致命影响。这些报告作为第一个铭文，依次与另一个铭文相关联，然后是另一个铭文，直到最终它在程序源代码中指示现象的一个起源。此外，对齐实践也可以自动化并集成到编程语言本身中。通常，解释器或编译器会自行指示其断点，即对其轨迹产生负面影响的脚本行。但是，如果这些错误报告对程序员来说似乎是自动的，那么不应忘记它们是异质过程的产物，因为参与编程语言维护和增强的编程团队必须应对铭文的对齐，以便首先确定应该索引哪种类型的错误。8^虽然在扩展和所涉及的工作量方面有所不同，但这些程序测试、错误报告和编程语言设计过程也可能与对齐铭文和生成参考链有关。

For the case of computer programming, one may imagine different expressions of the alignment practices I have documented above. Even though I conjecture that these expressions still consist in forming chains of reference in order to access remote entities and point at specific locations within numbered lists of inscriptions, they may not necessary deploy themselves in a spatio-temporal landmark that is similar to the one of DF. If we consider for example “program testing”—an important industrial process that consists in detecting and documenting errors in order to modify lines of code—this work can be highly distributed in space and time (Parrington and Roper 1989; Myers, Sandler, and Badgett 2011).⁷ The “bug reports” we often encounter when one of our software programs crash for mysterious reasons are other expressions of this necessity to align inscriptions because they consist precisely in documenting at what time and following what actions the program fatally affected the interpreter, compiler, or processor. These reports serve as first inscriptions that will, in turn, be articulated with another one, and then another one, until eventually it indicates one origin of the phenomenon within the source code of the program. Moreover, alignment practices can also be automated and integrated within the programming languages themselves. This is typically the case when an interpreter or compiler indicates by itself its breakpoint, the line of the script that negatively affects its trajectory. But if these error reports appear automatic to the programmer, it should not be forgotten that they are the product of heteromatic processes as the programming teams involved in the maintenance and enhancement of programming languages have to cope with alignment of inscriptions in order to establish what type of errors should be indexed in the first place.⁸ While different in terms of extension and labor involved, these processes of program testing, bug reporting, and programming language design are also, possibly, about aligning inscriptions and producing chains of reference.

对齐铭文以识别书面符号编号列表中的位置的做法也可以至少部分解释专业程序员对程序可理解性的痴迷。9 ^Button和 Sharrock (1995) 在他们令人钦佩但孤独的计算机编程实践研究中很好地记录了这一主题。正如他们所表明的那样，让其他程序员理解程序涉及传统的命名变量和函数，使其结构可读性更强，就像一个有组织的参考文档。它还涉及格式化和布局代码的不同函数和参数，使其易于从其视觉组织中浏览。这通常还包括通过小的解释性句子对程序进行注释，其首字母符号（“％” (以 Matlab 为例) 允许解释器或编译器忽略它们。如果我们刚刚遵循的编程序列不直接处理格式化、布局和注释，它仍然指定了这些实践所追求的目标。鉴于上面提出的要素，命名、格式化和注释都指向未来的时刻，它们可以作为直接可登记在引用链构成中的地标。因此，这些标记可以形成一个额外的参考基础设施，能够在解释器或编译器未来受到负面影响的情况下加速对齐工作（这很可能发生在需要维护和增强复杂程序的公司环境中）。

The practice of aligning inscriptions to identify locations within numbered lists of written symbols may also explain, at least in part, the obsession of professional programmers with program intelligibility.⁹ This topic has been well documented by Button and Sharrock (1995) in their admirable, yet solitary, study of computer programming practices. As they showed, making a program intelligible to other programmers involves conventional naming of variables and functions to make its structure readable as an organized and referenced document. It also involves formatting and laying out the different functions and parameters of the code to make it easily browsable from its visual organization. This also typically includes commenting on the program by means of small explicative sentences whose initial symbols (“%” for the case of Matlab) allow them to be ignored by interpreters or compilers. If the programming sequence we have just been following does not directly deal with formatting, laying out, and commenting, it nonetheless specifies what these practices are striving toward. In view of the elements presented above, naming, formatting, and commenting all point to future moments when they can operate as landmarks directly enrollable in the constitution of chains of reference. These marks may thus form an additional referential infrastructure capable of accelerating alignment work in the event of a future negative affection of an interpreter or a compiler (which is likely to happen in corporate settings where complex programs have to be maintained and enhanced).

但是，计算机编程的对齐实践是否等同于科学中的实验室实践？当然不是，现在是时候介绍它们之间的重要区别了。编程的对齐实践导致在脚本中识别位置，而科学实验室实践通常导致定义新对象，这些对象的属性和轮廓随后在学术论文中呈现并在同行之间讨论。当我们在第 5 章和第 6 章讨论数学时，我们将回到科学知识形成的这一关键方面。现在，可以说，计算机编程中对齐实践的动力和结果主要涉及试图完成足够脚本的程序员，而科学实验室中的对齐实践则转向完成有说服力的书面声明。科学实验室始终是反实验室（Latour 1987，79-100）：它们也应被理解为一种发表比竞争对手更有力的主张的手段。科学实验室实践的竞争性方面不断试图确定什么应该算作自然，因此必须将其与计算机编程实验室实践的自指性方面区分开来：科学家试图证明他们实际制造的现象的客观现实，而程序员则试图遵循他们所依赖的场景（稍后会详细介绍）。简而言之，科学家和程序员所处的网络我认为，参与的程度是相当不同的。科学领域的铭文对齐支持声明的发表，而计算机编程领域的对齐实践则支持技术产品的完成，但该产品在企业环境中仍需易于理解。

But are the alignment practices of computer programming equivalent to the laboratory practices in the sciences? Of course not, and it is now time to present an important difference between them. Whereas the alignment practices of programming lead to the identification of a location within a script, scientific laboratory practices generally lead to the definition of new objects whose properties and contours are later presented in academic papers and discussed among peers. We will come back to this crucial aspect of the formation of scientific knowledge when we will consider mathematics in chapters 5 and 6. For now, suffice it to say that whereas both impetuses and outcomes of alignment practices in computer programming mainly concern programmers who try to complete adequate scripts, alignment practices in scientific laboratories are turned toward the completion of persuasive written claims. Scientific laboratories are always counter-laboratories (Latour 1987, 79–100): they are also to be understood as a means to publish stronger claims than their competitors. The agonistic aspect of laboratory practices in the sciences that constantly try to establish what should count as natural must then be demarcated from the self-referential aspect of laboratory practices in computer programming: While scientists try to make a case for the objective reality of the phenomena they practically make appear, programmers try to follow a scenario they are attached to (more on this later). In short, the networks in which scientists and programmers participate are, I believe, quite dissimilar. Whereas alignment of inscriptions in the sciences support the publication of claims, alignment practices in computer programming support the completion of a technical artifact that yet needs to be intelligible in corporate settings.

因此，科学实践与编程实践之间的相似性有其局限性。但我相信这两种实践都具有一些至关重要的（而且相当令人惊讶的）相似性，它们都允许形成参考链并访问远程实体。就像科学工作一样，计算机编程不能简化为这种特定类型的实践。事实上，一旦到达远程实体，一旦确定了有问题的位置，仍然需要进行许多操作。在这方面，对齐铭文只是编程活动的一小部分。

The analogy between scientific and programming practices therefore has its limits. Yet I also believe that both practices share some crucial—and quite surprising—similarities, both allowing the formation of chains of reference and access to remote beings. And just like scientific work, computer programming cannot be reduced to this specific type of practice. Indeed, once the remote entity has been reached, once the problematic location has been localized, many operations still need to be conducted. In this respect, aligning inscriptions is only a small part of the activity of programming.

技术弯路

Technical Detours

我们在上一节中看到，有时在编程过程中，当一个小的、快速的、难以掌握的实体（例如，解释器、编译器、微处理器）在其轨迹上受到影响，以至于无法再触发电脉冲来计算数据时，程序员需要增加铭文、对齐它们并将它们堆积起来，直到铭文构成对实体的某种访问 - 反过来，这种访问会指示脚本中的某个位置。但接下来会发生什么？

We saw in the previous section that sometimes, during programming episodes, when a small, swift, and difficult-to-grasp entity (e.g., an interpreter, a compiler, a microprocessor) is affected in its trajectory to the point of not being able to trigger electric pulses for the computation of data anymore, programmers need to multiply inscriptions, align them, and pile them up until the inscriptions constitute some access to the entity—access that, in turn, indicates a location within the script. But what happens next?

在本节中，我们将重点介绍在编程过程中部署的另一组实践。虽然这组实践肯定与铭文的对齐有关，但它具有不同的含义。编程的科学方面涉及添加和对齐铭文（实验、确认、“愚蠢的事情”）以到达远程实体，而我称之为编程的技术方面则涉及包含和替换实体以绕过僵局。再一次，希望这句奇怪的句子会随着章节的继续而变得更加清晰。现在，我们将继续遵循 PROG，从上一个序列结束的地方开始：

In this section, we will focus on another set of practices deployed during programming episodes. While this set of practices surely goes along the alignment of inscriptions, it has different implications. Whereas the scientific aspect of programming involves the addition and alignment of inscriptions (experiments, confirmations, “stupid things”) in order to reach a remote entity, what I shall call the technical aspect of programming involves the inclusion and substitution of entities to get around impasses. Once again, this odd sentence will hopefully become clearer as the chapter goes on. For now, we shall continue to follow PROG, starting exactly when the previous sequence ended:

[DF检查图4.18 —T8的命令窗口]

[DF examines the Command Window of figure 4.18—T8]

国防军：  “它是这个 'j+3' 在第一个矩形之后，它变得太大了。它采用第一个矩形，如果第二个矩形更大，它就无法增加。所以我只是按某种顺序排列。”

DF:  “It is this ‘j+3’ that becomes too big after the first rectangle. It takes the first rectangle and if the second rectangle is bigger, it just can’t increment. So I’ll just put in some order.”

[DF删除第3、8和10行；删除“；”在第 9 行末尾]

[DF deletes lines 3, 8 and 10; deletes “;” at the end of line 9]

[DF运行脚本]

[DF runs the script]

[图4.19 —T9]

[figure 4.19—T9]

国防军：  “好的，我们只需要改变一些东西。”

DF:  “OK, we just need to change a few things.”

正如我们在上一节中看到的，DF 成功地定位了严重影响 INT 的脚本行。为了建立这种经过认证的知识，必须制作和对齐几个铭文。但这些铭文现在毫无用处；它们只是作为 DF 对 INT 的准科学探究的一部分而相关。现在是时候让 DF 真正改变脚本中的一些内容了。为此，他首先整理了一些内容并删除了用作实验工具的说明（图 4.19）。

As we saw in the previous section, DF managed to localize the line of the script that is badly affecting INT. Several inscriptions had to be produced and aligned in order to establish this certified knowledge. But these inscriptions are now useless; they were only relevant as part of DF’s quasi-scientific inquiry into INT. It is now time for DF to really change a few things in the script. To do so, he starts by putting in some order and deleting the instructions that were used to him as experimental instruments (figure 4.19).

在本章的这个阶段，为了解释接下来发生的事情，我需要引入一个补充符号，以便我们更好地掌握DF 即将进行的技术创新。根据技术项目的历史和社会学研究的结果，我将借鉴的符号是在 20 世纪 90 年代提出的，旨在说明技术项目的演变，而不使用传统的、有问题的自然与社会之间的区别（Latour、Mauguin 和 Teil1992 年）。我们不需要了解这种映射的所有微妙之处，顺便说一句，这种映射从未真正流行起来。10^对于我们感兴趣的内容，我们将仅介绍这些所谓的社会技术图（STG）的基本原理。

At this point of the chapter, to account for what happens next, I need to introduce a complementary notation that will allow us to have a better grip on the technical innovations DF is about to conduct. Following results of historical and sociological studies of technical projects, the notation I will draw on has been proposed during the 1990s as an attempt to illustrate the evolution of technical projects without using the traditional and problematic distinction between nature and society (Latour, Mauguin, and Teil 1992). We do not need to understand all the subtleties of this mapping that, by the way, never really took off.¹⁰ For what interests us here, we shall only cover the basic principles of these so-called sociotechnical graphs (STGs).

社会技术项目研究结果之一是，这些项目的发展轨迹取决于它们吸收新行动者（人类或非人类实体）以克服关键僵局的能力（Akrich 1989；Callon 1986；Latour 1993a）。历史上此类吸收的例子不胜枚举：为了使美国贝尔公司在美国电话网络开发中胜过西联汇款公司，它不得不在经过多次诉讼后，在其社会技术网络中吸收关键电话专利（Brooks 1976）。通过实现高可靠性和灵活性开关晶体管的生产，平面工艺使仙童半导体公司成为美国空军的商业合作伙伴（L é cuyer、Brock 和 Last 2010）。通过引入 20 世纪 60 年代初麻省理工学院开发的分时技术，John Kemeny 及其团队得以在达特茅斯学院继续开发 BASIC 编程语言（Montfort 等人，2013，158–194）。对于每个示例，都引入了一个特定的行为因素（一组电话专利、平面过程、分时技术），这反过来又使项目略有变化。技术历史和社会学的一个重要功劳是相继证明了引入新的行为因素对于技术项目发展的重要性——它们可能像 19 世纪末美国的电气化一样巨大（Hughes 1983；Nye 1992），也可能像安装道路减速带一样小（Latour 2006）。

One of the results of the studies of sociotechnical projects was to show that the trajectories of such projects are a function of their capacity to enroll new actants—human or nonhuman entities—in order to overcome critical impasses (Akrich 1989; Callon 1986; Latour 1993a). Historical examples of such enrollments are legion: in order for American Bell to prevail over Western Union in the development of the telephone network in the United States, it had to enroll—after many lawsuits—crucial telephone patents within its sociotechnical network (Brooks 1976). By enabling the production of highly reliable and flexible switching transistors, the planar process allowed Fairchild Semiconductor to become a commercial partner of the US Air Force (Lécuyer, Brock, and Last 2010). By enrolling the time-sharing technology as developed at MIT at the beginning of the 1960s, John Kemeny and his team were able to pursue the development of the BASIC programming language at Dartmouth College (Montfort et al. 2013, 158–194). For each example, a specific actant—a set of telephone patents, the planar process, the time-sharing technology—is enrolled, and this, in turn, makes the project slightly shift. One important credit to the history and sociology of technologies is to have successively demonstrated how crucial the inclusion of new actants for the development of technical projects is—may they be huge as the electrification of the United States at the end of the nineteenth century (Hughes 1983; Nye 1992) or small as the installation of a road bump (Latour 2006).

然而，技术项目吸收新参与者以进行发展和扩展的“纬度”维度如果没有一个正交的“纵向”维度来表达新参与者所建议的转变，那么它就是不完整的。技术项目的历史和社会学的另一个重要结果是，新参与者的加入同时改变了项目先前参与者之间的关系，从而可能造成新的僵局。用上一段的例子来说，贝尔的技术系统因电话专利的加入而发生了变化：这个之前很小的网络成为美国电话通信的潜在垄断者，因此需要进一步重新配置，以免成为美国司法部反垄断诉讼的目标（Gertner 2013）。仙童半导体公司因平面工艺的加入而发生了根本性的转变：它成为一个强大的实体很快就能够进行集成电路的工业生产。这些生产能力反过来又参与了洲际弹道导弹的开发，这进一步引发了对集成电路的需求激增，并逐渐形成了强大的竞争对手（最著名的是德州仪器和摩托罗拉；见 Campbell-Kelly 等人，2013 年，第 210-225 页）。同样，达特茅斯计算机系统纳入分时技术也极大地参与了 BASIC 编程语言的设计，大大增加了其 beta 测试。但行动者“分时”的加入也改变了达特茅斯的计算基础设施，通过允许学生广泛使用，它很快就开始用于原创的计算机游戏实验（Montfort 等人，2013 年，第 165-194 页）。除了招收（或失去）行动者之外，技术项目也由他们修改。而且，就像技术项目的纬度（包容性）轴一样，这个经度（变革性）轴不仅涉及大型且高度复杂的技术系统：小型平凡的项目也会受到它的影响（Latour 1992）。

Yet, this “latitudinal” dimension of technical projects enrolling new actants in order to develop and expand would be incomplete without an orthogonal “longitudinal” dimension expressing the transformations suggested by the newly enrolled actants. Another crucial result of the history and sociology of technical projects is indeed that the inclusion of new actants simultaneously modifies the relationships among the previous actants of the project, thus potentially creating new impasses. Using the examples of the previous paragraph, Bell’s technical system was transformed by the inclusion of telephone patents: the previously tiny network became a potential monopoly over telephone communications in the United States, hence necessitating further reconfigurations so as not to be the target of antitrust lawsuits by the US Department of Justice (Gertner 2013). Fairchild Semiconductor was fundamentally transformed by the inclusion of the planar process: it became a powerful entity soon capable of industrial production of integrated circuits. These production capacities participated, in turn, in the development of intercontinental ballistic missiles, and this further created an explosion of the demand for integrated circuits and the progressive formation of serious competitors (most notably, Texas Instruments and Motorola; see Campbell-Kelly et al. 2013, 210–225). Similarly, the inclusion of time-sharing technology within Dartmouth’s computer system greatly participated in the design of the BASIC programming language by considerably increasing its beta testing. But the inclusion of the actant “time sharing” also transformed Dartmouth’s computing infrastructure, which, by allowing its extensive utilization by students, soon started to be used for original computer-game experiments (Montfort et al. 2013, 165–194). More than just enrolling (or losing) actants, technical projects are also modified by them. And just like the latitude—inclusive—axis of technical projects, this longitude—transformative—axis does not only concern large and highly complex technological systems: small mundane projects are also affected by it (Latour 1992).

基于技术项目的这种双重性以及从语言学中借用的概念，STG 的支持者提出了一种根据两个维度来映射技术项目发展的方法：组合维度和聚合维度。STG 的第一个维度（组合）由特定时间 T 时行为元的特定组合定义。行为元在时间 T 的配置特定于每个技术项目，因此应通过揭示正在考虑的项目的原因和理由的叙述来支持。由于这个维度表达了变量之间的关联，因此可以将其称为 AND 维度。AND 维度中行为元的配置分为两个分支：配置参与项目发展的“盟友”和配置对项目完成构成障碍的“对手”。同样，哪些行动者被视为项目发展的盟友或对手取决于 STG 所概括的叙述（Latour、Mauguin 和 Teil 1992，39）。区分盟友行动者配置和对手行动者配置的边界构成了时间 T 时技术项目的“前线”。

Building on this dual aspect of technical projects as well as concepts borrowed from linguistics, the proponents of STGs proposed a way to map the development of technical projects according to two dimensions: a syntagmatic dimension and a paradigmatic dimension. The first dimension (syntagmatic) of STG is defined by specific assemblages of actants at a certain time T. This configuration of actants at a time T is specific to each technical project and should therefore be supported by a narrative that exposes the whys and wherefores of the project being considered. As this dimension expresses association among variables, it can be called the AND dimension. The configuration of actants in the AND dimension is separated into two branches: the “allies” whose configuration participates in the development of the project and the “opponents” whose configuration constitutes an obstacle to the completion of the project. Again, which actant counts as an ally or as an opponent to the development of the project depends on the narrative the STG is only summarizing (Latour, Mauguin, and Teil 1992, 39). The boundary that separates allies’ configuration of actants and opponents’ configuration of actants constitutes the “frontline” of the technical project at time T.

第二个维度（典型维度，与托马斯·库恩的概念无关）由在时间 T + 1 时发生在盟友和对手配置中的替换定义。由于这个维度表达了变量的替换，因此可以将其称为 OR 维度。取决于T + 1时盟友和对手配置的波动，技术项目的前线也可能波动。同样，哪个行动者被另一个行动者取代，从而可能使前线波动，取决于技术项目的叙述。

The second (paradigmatic; nothing to do with Thomas Kuhn’s notion) dimension is defined by the substitutions that have occurred in both allies’ and opponents’ configurations at time T + 1. Since this dimension expresses substitution of variables, it can be called the OR dimension. Depending on the fluctuation of allies’ and opponents’ configurations at T + 1, the frontline of the technical project may also fluctuate. Once again, which actant is substituted by another, thus potentially making the frontline fluctuate, depends on the narrative of the technical project.

将技术项目叙述转化为 STG 还需要另外两个元素：指定的观点和我所说的“场景”。首先，必须指定 STG 总结的行动者对项目的看法。从这个意义上讲，对于任何给定的技术项目叙述，如果该叙述采用了许多不同行动者的观点，则每个观点都可以（潜在地）由一个特定的 STG 来映射。其次，还必须指定被映射观点的行动者的愿望。这个话题很棘手，将在本章的下一节中进一步展开。现在，可以说，每个STG 都应该指定行动者想要实现的目标、它想要生活的未来以及它所依附的场景。

Two other elements are necessary to translate the narrative of a technical project into an STG: a specified point of view and what I call a “scenario.” First, the point of view of the actant whose view on the project is being summarized by the STG has to be specified. In that sense, for any given narrative about a technical project, if this narrative takes the point of view of many different actants, each point of view can (potentially) be mapped by one specific STG. Second, the desire of the actant whose point of view is being mapped also has to be specified. This topic is a tricky one and will be further developed in the next section of this chapter. For now, suffice it to say that what the actant wants to achieve, the future it wants to live in, the scenario to which it is attached should be specified in each STG.

现在让我们尝试将这些理论元素应用到我们感兴趣的项目中：DF 完成 PROG 的项目。如果我们考虑 T8 及其之前的整个叙述，我们可能能够将其转化为 STG，总结 DF 的盟友和对手。图表的第一个元素应该表明它所代表的观点。与大多数关于大型技术系统的叙述相反，在这些叙述中，许多观点被考虑和对抗，而我们的小叙述只考虑了 DF 的观点。图表的第二个元素应该是DF 所附加的场景。正如上一节已经提到的，我们知道 DF 的 PROG 场景可以总结如下：“创建一个矩阵，其像素值对应于工人在每个像素上绘制的矩形数量。”关于执行者：脚本的每条指令都可以被视为一个执行者，因为它们都让 INT 做事。但其他执行者也可能包含在图表中，只要它们对场景框架中的项目有影响。从这个意义上讲，命令行中打印的红色铭文以及这些铭文根据 DF 所指的内容以及脚本打算对 .txt 文件的数据执行的最终操作也可以包含在 STG 中。此外，由于脚本项目的叙述表明现在多个指令已经稳定，我们可以将这些“稳定的指令包”视为一个单一的执行者。如果我们将这些元素放在一起考虑并使其适应 T8，我们最终会得到一个如图 4.20所示的图表。

Let us now try to adapt these theoretical elements to the project that interests us here: DF’s project to complete PROG. If we consider T8 and the whole narrative that precedes it, we might be able to translate it into an STG summarizing DF’s allies and opponents. The first element of the graph should indicate the point of view that it re-presents. Contrary to most narratives about large technical systems where many points of view are considered and confronted, our small narrative only accounts for the point of view of DF. The second element of the graph should be the scenario to which DF is attached. As already touched upon in the previous section, we know that DF’s scenario for PROG can be summarized as such: “Creating a matrix whose pixel-values correspond to the numbers of rectangles drawn by workers on each pixel.” Concerning the actants: every instruction of the script can be considered an actant as they all make INT do things. But other actants might also be included in the graph as long as they impact on the project as framed by its scenario. In that sense, the red inscriptions printed in the Command line and what these inscriptions refer to according to DF as well as the final actions the script is intended to accomplish on the data of the .txt file can also be included in the STG. Moreover, as the narrative of the script-project indicates that several instructions are now stabilized, we may consider these “stable packages” of instructions as one single actant. If we consider these elements altogether and adapt them for T8, we end up with a diagram that looks like figure 4.20.

重要的是要记住，T8 的 STG 映射是 T8 的简化，最初以 Matlab 视图呈现，并由 DF 的说法丰富。与任何简化一样，它省略了许多元素。但与许多简化一样，它也可以作为一种工具来识别混乱过程的关键特征（Star 1983）。

It is important to remember that the STG mapping of T8 is a simplification of T8 as initially presented in its Matlab view and enriched by DF’s sayings. As any simplification, it omits many elements. But as many simplifications, it may also work as an instrument to identify key features of messy processes (Star 1983).

从这一点来看，基于上面的叙述，我们可以将T9纳入STG图，从而稍微修改盟友和对手的配置（见图4.21）。

From this point, based on the narrative presented above, we can include T9 in the STG graph, thus slightly modifying the configurations of allies and opponents (see figure 4.21).

对于此编程序列中剩余的每个 T，我将首先提供其完整叙述（简化的 Matlab IDE 和 DF 语录的转录），简要讨论它，然后提供其 STG 翻译。由于“观点”和“场景”在整个编程序列中都不会改变，因此从现在开始我将忽略它们。此外，在每个新的 STG 中，我都会用粗体突出显示新注册的执行者。在编程序列的最后，当 DF 完成 PROG 时，所有 STG 的连续性应该使我们能够检测到程序员部署的另一组实践，这些实践与铭文的对齐相一致，但我认为它们从根本上是不同的。

For each remaining T of this programming sequence, I will first present its complete narrative (simplified Matlab IDE and transcriptions of DF’s sayings), discuss it shortly, and then present its STG translation. As both “point of view” and “scenario” will not change throughout the programming sequence, I will ignore them from now on. Moreover, in every new STG, I shall highlight the newly enrolled actant in bold. At the very end of the programming sequence, when DF will have completed PROG, the succession of all the STGs should allow us to detect another set of practices deployed by programmers that goes along with the alignment of inscriptions while being, I believe, fundamentally different.

让我们继续关注 DF，看看他如何努力塑造 PROG：

Let us continue to follow DF as he tries to shape PROG:

国防军：  “我们会这样做。”

DF:  “We’re gonna do it like this.”

[DF 在 7 处创建新行]

[DF creates a new line at 7]

国防军：  “如果 'j+3' 更大”

DF:  “If ‘j+3’ is larger”

[在第 7 行，输入“如果 j+3 >”]

[at line 7, types “if j+3 >”]

国防军：  “比用户细胞的大小”

DF:  “than the size of the cell of the user”

[在第 7 行，输入“大小（用户{j}）”]

[at line 7, types “size(users{j})”]

国防军：  “然后它就过去了”

DF:  “then it goes over it”

[DF 在 8 处创建新行；输入“休息”]

[DF creates a new line at 8; types “break”]

[DF运行脚本]

[DF runs the script]

[图4.22 —T10]

[figure 4.22—T10]

[DF检查图4.22 —T10的命令窗口]

[DF examines Command Window of figure 4.22—T10]

国防军：  “啊，当然了。我不应该接受'杰“无法用这种方式定义任何事。”

DF:  “Argh, of course. I shouldn’t take ‘j.’ Can’t define anything that way.”

在 T10 时（图 4.22），DF 注册了一个新的执行者：“如果”语句，从第 7 行开始，到第 9 行结束。因为此时，他知道如果第二个矩形大于第一个矩形，INT 将被阻止，因此添加一个条件语句，要求 INT 解决这个维度问题，是完全合理的。添加一个“如果”语句将允许 INT 继续解释脚本，即使它遇到比前一个更大的矩形。但是，由于红色铭文和 DF 的表示，该声明表达不当：DF不应该采取“杰”作为“的大小变量用户”因为它在第 5 行就已经等于零了。这个归因错误导致的后果是 INT无法定义任何东西。无法定义任何矩形，因此矩阵也无法递增。

At T10 (figure 4.22), DF enrolls a new actant: the “if” statement that starts at line 7 and ends at line 9. Since, at this point, he knows for a fact that INT is blocked if the second rectangle is bigger than the first one, the addition of a conditional statement that could ask INT to go over this dimension problem makes complete sense. The addition of an “if” statement would thus allow INT to continue its interpretation of the script even though it encounters a rectangle bigger than the previous one. But as the red inscription and DF’s saying indicate, the statement was inappropriately expressed: DF should not have taken “j” as the size variable of “users” since it already equals to zero at line 5. The consequence of this attribution mistake is that INT cannot define anything. No rectangle can be defined, and the matrix cannot, in turn, be incremented.

如果我们将 T10 映射为与 T8 和 T9 一致的 STG，我们将获得图 4.23。查看它，我们可以看到出现了新的执行者并给项目带来了差异，稍微改变了它的战线。在盟友的配置中，DF 添加了“I”，以绕过“W”、“X”、“Y”和“Z”的配置。但是，如果这个新的执行者让“W”和“X”消失（即索引不再超过矩阵维度，并且第二个矩形不再太大），那只有通过出现两个新的对手：“V”和“U”。那么“Y”和“Z”仍然坚决抵制 DF 的项目，因为此时无法定义任何矩形。

If we map T10 as an STG in line with T8 and T9, we obtain figure 4.23. Looking at it, we can see that new actants have appeared and created differences in the project, slightly altering its frontline. In the allies’ configuration, “I” has been added by DF in order to get around the configuration of “W,” “X,” “Y,” and “Z.” But if this new actant made “W” and “X” disappear—that is, the index does not exceed the matrix dimension anymore, and the second rectangle is not too big anymore—it is only by making two new opponents appear: “V” and “U.” “Y” and “Z” are then still solidly opposing resistance to DF’s project since, at this point, no rectangle can be defined.

让我们继续：

Let us continue:

[在第 7 行，DF 删除“用户，{j}）”]

[at line 7, DF deletes “users, {j})”]

国防军：“实际上，细胞的大小应该只是‘用户，2'”

DF: “Actually, the size of the cell should just be ‘users, 2’ ”

[在第 7 行，输入“用户,2)”]

[at line 7, types “users,2)”]

[运行脚本]，

[runs the script],

[图4.24 —T11]

[figure 4.24—T11]

国防军：  “好的，也许有用。”

DF:  “OK, it may work.”

在 T11 处（图 4.24），DF 修改了条件指令：不再引用“杰”新矩形的大小现在指的是单元格的第二个值，“用户”。我们不需要准确理解这个值和单元格指的是什么。T11 中重要的是，一个新的动作元素的加入——修改后的“如果”语句——产生了一个重要的区别：INT 不再打印红色铭文。这表明 INT 已成功翻译每一行，从而触发了对 .txt 文件数据的电子计算。此时，它可能会起作用：矩形可能会增加空矩阵。但它还没有结束，因为对称地，它也可能不起作用。由于命令窗口没有提供有关空矩阵增加的任何指示，因此可能还发生了其他事情。

At T11 (figure 4.24), DF modifies the conditional instruction: instead of referring to “j,” the size of the new rectangle now refers to the second value of the cell, “users.” We do not need to understand precisely what this value and cell refer to. The important thing at T11 is that the inclusion of a new actant—the modified “if” statement—creates an important difference: INT does not print a red inscription anymore. This indicates that INT has managed to translate every line, thus triggering electronic computation on the data of the .txt files. At this point, then, it may work: the rectangles may increment the empty matrix. But it is not over yet since, symmetrically, it may also not work. Since the Command Window does not provide any indication about the incrementation of the empty matrix, something else may also have happened.

如果我们通过包含 T11 继续我们的 STG 重新表示这个编程序列，我们将得到图 4.25。几项变化影响了 T11 时盟友的配置。“I”消失了：DF 删除了它，因为它只有通过出现新的对手才能使对手消失。但包含了两个新的动作：“J”对应于新的条件语句，“K”对应于命令窗口中没有任何错误铭文（以及 DF 的断言“它可能有效”的推论）。这种新的盟友配置是否设法绕过了对手的配置？只是部分如此，因为“K”所暗示的不确定性有其推论：由于命令窗口中没有迹象表明脚本也不起作用（“T”），也就是说，它可能无法正确增加空矩阵。因此，“Z”（“脚本不符合场景”）保持不变。此时，DF 仍然需要包含其他内容；他仍然需要通过其他方式来推进他的项目，以绕过“ T”和“Z”构成的僵局。

If we continue our STG re-presentation of this programming sequence by including T11, we obtain figure 4.25. Several changes affected the allies’ configuration at T11. “I” disappeared: DF deleted it because it made opponents disappear only by making new ones appear. But two new actants are included: “J” that corresponds to the new conditional statement and “K” that corresponds to the absence of any error inscription within the Command Window (and, corollary, to DF’s assertion that “it may work”). Did this new configuration of allies managed to get around the configuration of opponents? Only partially since the incertitude suggested by “K” has its corollary: as there is no indication in the Command Window, the script may also not work (“T”), that is, it may not increment the empty matrix properly. As a consequence, “Z”—“the script does not follow the scenario”—holds on. At this point, DF still needs to include something else; he still needs to pursue his project by other means in order to get around the impasse constituted by “T” and “Z.”

让我们继续关注DF：

Let us continue to follow DF:

国防军：  “但我只是想确定一下。”

DF:  “But I just need to be sure.”

[创建第 20 行；输入“展示(R)”]

[creates a line 20; types “imshow(R)”]

[运行脚本]

[runs the script]

[图 4.26 —T12 和图 4.27 ]

[figure 4.26—T12 and figure 4.27]

缩略词：  “太接近了！”

FJ:  “This is close!”

国防军：  “是的。但它在值 1 之后被截断。”

DF:  “Yep. But it clips after the value 1.”

缩略词：  “剪辑？”

FJ:  “Clips?”

国防军：  “是的，它经常这样做。基本上，它不考虑任何大于 1 的值。我的意思是，矩阵可能有多个值，但它不会在图像上显示出来。”

DF:  “Yes, it often does that. Basically, it doesn’t consider anything above 1. I mean, the matrix may have values more than one, but it does not show it on the image.”

在T12（图4.26）处，DF添加了一条新指令—“展示(R)”——要求 INT 打印递增矩阵的图像（图 4.27）。结果既令人信服又令人失望。积极的一面是矩阵已经有效递增。INT 打印的图像证明了这一点：它具有不同的值，这些值共同形成一个白色形状。但消极的一面是，该图像只有二进制值：1 形成白色形状，0 形成黑色背景。根据 DF 的说法，INT 再次是这个问题的根源：通过在值 1 之后进行剪切，打印的图像只能是二进制的。在这些条件下，很难知道递增矩阵的构成值。此时，DF 再次需要在脚本中包含其他内容，以使其遵循所需的场景。

At T12 (figure 4.26), DF adds a new instruction—“imshow(R)”—that asks INT to print an image of the incremented matrix (figure 4.27). The results are convincing as well as disappointing. The positive thing is that a matrix has effectively been incremented. The image printed by INT attests to this: it has differentiated values that together form a white shape. But the negative thing is that this image has only binary values: ones forming the white shape and zeros forming the black background. According to DF, INT is once again the cause of this problem: by clipping after the value 1 the printed image can only be binary. In these conditions, it is difficult to know what values constitute the incremented matrix. At this point, again, DF needs to include something else in the script in order to make it follow the desired scenario.

让我们看一下 STG，简要了解一下刚刚发生的事情（图 4.28）。盟友的配置再次扩大：“L”和“M”允许 DF 确保矩形增加矩阵。这反过来又使“T”消失，因此该项目的这一方面不再存在不确定性。但“M”的二元特性使“R”出现在对手的配置中：由于未知原因，INT在第一之后被剪辑。这反过来又产生了“S”，即对脚本递增能力的不确定性，这种不确定性可能在“1在这些情况下，Z 仍然存在，并且脚本仍然没有遵循所需的场景。DF 再次别无选择：他必须在盟友配置中注册其他内容；他必须将“R”、“S”和“Z”的变通方法委托给新的执行者。

Let us have a look on the STG to get a condensed look on what has just happened (figure 4.28). The configuration of allies has again expanded: “L” and “M” allowed DF to be sure that the rectangles increment the matrix. This, in turn, made “T” disappear so that no incertitude remains concerning this aspect of the project. But the binary characteristic of “M” made “R” appear in the configuration of opponents: for unknown reasons, INT clips after one. This, in turn, creates “S,” the incertitude about the incrementing capability of the script that may stop after “1.” In these conditions, Z remains, and the script is still not following the desired scenario. Once again, DF has no choice: he has to enroll something else to the configuration of allies; he has to delegate the work-around of “R,” “S,” and “Z” to a new actant.

考虑到这些因素，让我们继续：

With these elements in mind, let us continue:

国防军：  “所以我会尝试除以'R' 除矩阵的最大值外。如果它有其他值，则应该显示出来。”

DF:  “So I’ll just try to divide the value of ‘R’ by the maximal value of the matrix. If it has other values than one, it should show it.”

[在第 20 行，输入“/最大(R(:))”]

[at line 20, types “/max(R(:))”]

[运行脚本]

[runs the script]

[图 4.29 —T13 和图 4.30 ]

[figure 4.29—T13 and figure 4.30]

国防军：  “好的，这是矩阵的正确图像。就是这个。”

DF:  “All right, this is the right image of the matrix. This is it.”

通过包含这最后一小段代码——”/最大(R(:))”——DF 成功完成了脚本（图 4.29）。没有任何不确定性：矩阵正确地递增，如新输出图像所示（图 4.30）。因此，DF 成功地让 INT 根据宽度和高度值设计一个空矩阵；根据宽度、高度和位置值定义矩形；并使用这些矩形连续增加矩阵的像素值。虽然必须进行几项技术操作，但最终，该项目实现了其最初的目标。此时，可以将脚本视为执行可定义操作的技术工件。

By including this last small bit of code—“/max(R(:))”—DF manages to complete the script (figure 4.29). No incertitude remains: the matrix is correctly incremented as the new output image shows (figure 4.30). DF thus successfully managed to make INT design an empty matrix according to width and height values; define rectangles from width, height, and position values; and use these rectangles to successively increment the pixel-values of the matrix. Several technical operations had to be conducted but, in the end, the project fulfilled its initial ambitions. At this point, the script can be considered a technical artifact that does something definable.

如果我们看一下 STG（图 4.31），我们会发现“/最大(R(:))”成功地绕过了之前由“R”、“S”和“Z”形成的僵局。在 T13 中，“O”及其推论“P”的加入使得“R”、“S”和“Z”消失。指令“/最大(R(:))“ 制成INT 打印了矩阵的灰度图像，从而向 DF 展示其值确实在零和众包工人绘制的矩形总数之间变化。项目的所有反对者都已被盟友取代；所有死胡同都已被绕过。接下来是场景。正如 DF 所说，“就是这样。”编程序列结束了。

If we take a look at the STG (figure 4.31), we see that the inclusion of “/max(R(:))” managed to get around the impasse previously formed by “R,” “S,” and “Z.” At T13, the inclusion of “O” and its corollary “P” made “R,” “S,” and “Z” disappear. The addition of the instruction “/max(R(:))” made INT print a gray-scale image of the matrix, hence showing DF that its values do indeed variate between zero and the total number of rectangles drawn by the crowdworkers. All the opponents to the project have been replaced by allies; all dead-ends have been bypassed. The scenario is followed. As DF puts it, “this is it.” The programming sequence is over.

这些 STG 对我们分析这个编程序列有什么帮助？这种简化让我们看到了什么？虽然上一节强调了编程实践的科学性，但我认为本节强调的是编程实践的技术性。科学实践和技术实践不同吗？在行动的中间，它们肯定重叠到看似相似的程度。但是，按照 Latour (2013) 的说法，我仍然认为两者的表达方式截然不同。

What do these STGs add to our analysis of this programming sequence? What does this simplification allow us to see? While the previous section put the emphasis on the scientific moment of programming practices, I assume that the present section puts the emphasis on the technical moment of programming practices. Are scientific and technical practices different? In the middle of the action, they surely overlap to the point of appearing similar. But, following Latour (2013), I nonetheless assume that both express themselves quite differently.

我们看到，实验室科学实践与编程实践之间惊人的相似之处在于，它们都通过复制和对齐铭文来形成参考链，从而收集有关远程实体的信息。尽管这两项活动不能被视为等同，但我相信它们彼此呼应：有时两者都会生成和对齐铭文，以便访问远程实体。

We saw that the surprising similitude between the laboratory practices of science and the practices of programming lies in that they both multiply and align inscriptions in order to shape chains of reference, thereby allowing the assemblage of information about remote entities. Even though both activities cannot be considered equivalent, I believe they echo well with each other: both sometimes produce and align inscriptions in order to access remote beings.

尽管我们刚刚记录的序列需要形成（小）引用链才能启动，但我认为该序列还表达了一些完全不同的东西。在 T9 时，DF 需要更改脚本中的内容。他做了什么？在每个 T 时，他都会包括新的行为者并将行动委托给他们，以绕过阻碍场景执行的僵局。该序列中涉及的实践并不倾向于了解这些僵局；他们倾向于寻找绕过它们的方法。这正是 STG 最终成为指导工具的原因：通过简化叙述，它们允许遵循这些连续的转变，这种不断的曲折表达了新实体的注册、它们所暗示的委托以及它们触发的解决方法。脚本一旦在 T13 完成，就成为技术工件。但只有通过技术实践、巧妙的包含、委托和解决方法，这样的工件才能存在。结合完成的脚本，借助 STG 的简化，我们可以看一下 DF 绘制的雷击及其技术动作（图 4.32）。

Although the sequence we have just documented required the formation of a (small) chain of reference in order to be initiated, I assume the sequence also expressed something radically different. At T9, DF needed to change things in the script. What did he do? At each T, he included new actants and delegated actions to them in order to get around impasses that were obstructing the following of the scenario. The practices involved in this sequence did not tend toward gaining knowledge about these impasses; they tended toward finding ways to get around them. This is precisely why STGs were, in the end, instructive tools: by simplifying the narrative, they allowed to follow these successive shifts, this constant zigzag that expressed the enrollment of new entities, the delegation they implied, and the work-arounds they triggered. The script, once completed at T13, became a technical artifact. But it was only through technical practices, ingenious inclusions, delegations, and work-arounds that such an artifact could come to existence. Along with the finished script, thanks to the simplification provided by the STGs, we can glance at the lightning strike drawn by DF and its technical actions (figure 4.32).

该序列不是线性的；它由连续性的中断来安排节奏，而这些中断在脚本完成后就消失了。正如引用链在允许构成有关远程存在的信息后就被忽略一样，技术实践的不断转变、包含、委托和变通在允许完成工件后就变得不可见。我认为这是编程研究的一个严重局限性，因为编程研究只考虑编程测试的结果（见第 3 章）。通过只考虑最终的技术对象（完成的脚本），他们无法掌握该对象的技术性所必需的实践。只有从工件回溯到不断修改其形式从而使其变得独特的弯路，我们才能捕捉到计算机编程的技术方面。任何工作脚本都要感谢所有现在不可见的盟友，这些盟友被添加到每个配置中，以便绕过——甚至可以说，为了破解（Nissenbaum 2004）——现在同样不可见的对手。正如铭文的普及和排列使得 DF 变得知识丰富一样，技术上的迂回使他显得独辟蹊径：通过捕捉实体——精灵——并将它们纳入变通方案，他能够吸纳盟友并绕过对手，从而画出令人眼花缭乱的之字形。

The sequence was not linear; it was rhythmed by breaks of continuity that vanished at soon as the script was completed. Just as chains of reference are ignored as soon as they allowed the constitution of an information about a remote being, the constant shifts, inclusions, delegations, and work-arounds of technical practices are made invisible once they allowed the completion of the artifact. Here lies, I believe, a serious limitation of the studies of programming that only consider the results of programming tests (see chapter 3). By only considering the final technical object (the finished script), they cannot grasp the practices that were necessary to the technicality of this object. It is only by going backward from the artifact to the detours that have constantly modified its form, thus making it singular, that we may capture the technical aspect of computer programming. Any working script holds thanks to all the now-invisible allies that were added to each configuration in order to get around—one may even say, in order to hack (Nissenbaum 2004)—now also-invisible opponents. Just as the proliferation and alignment of inscriptions made DF become knowledge-able, the technical detours made him in-genious: by catching entities—jinns—and enrolling them in work-arounds, he was able to include allies and get around opponents, thus drawing a dazzling zigzag.

值得注意的是，当编程是关于绘制之字形时，这些类型的技术时刻往往是最受赞赏的时刻。虽然引用链的构建可能非常令人沮丧——铭文不断堆积而没有形成任何可靠的引用链——但绘制之字形所涉及的实践往往看起来更有趣。不幸的是，我无法通过任何经验材料支持这一说法；这意味着在本章的这一点上已经出现太多的其他编程形象。但在她对编程影响的文学描述之一中，艾伦·乌尔曼很好地表达了程序员在陷入技术弯路时经常体验到的这种感觉，一旦工件完成，就很难捕捉到：

It is interesting to note that these types of technical moments, when programming is about the drawing of a zigzag, are often the most appreciated ones. While the construction of chains of reference can be very frustrating—the inscriptions keep piling up without forming any reliable chain of reference—the practices involved in the drawing of zigzags often appears more playful. Unfortunately, I cannot support this claim by any empirical materials; this would imply the presentation of many other programming figures that are already too numerous at this point in the chapter. But in one of her literary accounts of programming affects, Ellen Ullman nicely expressed this feeling programmers often experience when they are engaged into technical detours that are very difficult to catch once the artifact is completed:

“该死！NULL 案例！”

“如果没有，我们就超出了文本字段，他们点击了空格——”

“——是的，就像——”

“—无参数—”

“地狱！”

“那我们太空飞行怎么样？”

“我不知道。……等一下！”

“是的，我们可以用太空垫——”

“—并将空格视为数字。”

“是的！我们将调用 SendKey(space) 来——”

“——数字对象。”

…

“不，不，不，不。如果集合的成员以空格开头怎么办？哦，天啊。”

他展现出的绝望近乎赤裸裸，这是我从未在电影之外见过的。在这里，在那个地方，我们没有羞耻。他看见过我睡在地板上，流着口水。我们都看见过丹尼在机房的高温下脱得只剩内衣，他那肿胀、发白的腹部——可惜他还年轻。我看见过乔尔的头皮屑，衣服上薄薄的一层猫毛，注意到过他身上我不该看到的东西。我敢肯定，他也看见过我黏糊糊的头发，注意到我没有化妆时看起来有多么暗淡，还看到过其他一些不便言及的私密细节。然而，这些都不再重要。我们的身体很久以前就被抛弃了，沦落到饥饿、失眠和连续几个小时坐在键盘和鼠标前的折磨中。我们的肉体已经被摧残殆尽。现在，我们只通过一种方式了解彼此：代码。

此外，我知道我现在可以给他带来一种在任何生命中都罕见的快乐：我即将把他从绝望中拯救出来。

“没问题，”我平静地说。我把手放在他的肩膀上，想做出一个安慰的手势。“参数从不以空格开头。”

正如我所希望的那样。他的绝望消失了。他变得充满活力，转向键盘，开始快速打字。现在他已经离开了我。他消失在代码中。（Ullman 2012，8–9；斜体添加）

“Damn! The NULL case!”

“And if not we’re out of the text field and they hit space—”

“—yeah, like for—”

“—no parameter—”

“Hell!”

“So what if we space-pad?”

“I don’t know. … Wait a minute!”

“Yeah, we could space-pad—”

“—and do space as numeric.”

“Yes! We’ll call SendKey(space) to—”

“—the numeric object.”

…

“No, no, no, no. What if the members of the set start with spaces. Oh, God.”

He is as near to naked despair as has ever been shown to me by anyone not in a film. Here, in that place, we have no shame. He has seen me sleeping on the floor, drooling. We have both seen Danny’s puffy, white midsection—young as he is, it’s a pity—when he stripped to his underwear in the heat of the machine room. I have seen Joel’s dandruff, light coating of cat fur on his clothes, noticed things about his body I should not. And I’m sure he’s seen my sticky hair, noticed how dull I look without make-up, caught sight of other details too intimate to mention. Still, none of this matters anymore. Our bodies were abandoned long ago, reduced to hunger and sleeplessness and the ravages of sitting for hours at a keyboard and a mouse. Our physical selves have been battered away. Now we know each other in one way and one way only: the code.

Besides, I know I can now give him pleasure of an order which is rare in any life: I am about to save him from despair.

“No problem,” I say evenly. I put my hand on his shoulder, intending a gesture of reassurance. “The parameters never start with a space.”

It is just as I hoped. His despair vanishes. He becomes electric, turns to the keyboard and begins to type at a rapid speed. Now he is gone from me. He is disappearing into the code. (Ullman 2012, 8–9; italics added)

在这个文学摘录中，一个信息逐渐被收集起来——叙述者写下了最后一句话（“参数从不以空格开头”）——然后一个位置被定义：让实体被注册，让动作被委托，让对手被绕过。而令人愉快的技术闪电很快就展开了。

In this literary excerpt, an information is progressively being assembled—the narrator provides the very last inscription (“The parameters never start with a space”)—and a location is, in turn, defined: let entities be enrolled, actions be delegated, and opponents be gotten around. And the joyful technical lightning strike soon unfolds.

如果我在本章的这个部分有点夸张，请读者原谅我，但这两节中记录的技术和科学实践为计算机编程提供了如此令人耳目一新的视角，我很难保持平静。我们确实看到，将计算机编程视为解决问题的过程的标准认知行为框架（参见第 3 章）可能会产生误导。程序员可能永远无法解决任何问题；当面对拒绝对数据产生电脉冲的远程实体时，他们或多或少共同构成了一个参考链，如果装备齐全，可能会指示一个有问题的位置，而这个位置反过来可能会触发新行为者的加入和僵局的技术解决方法。什么都没有解决；某个东西被定位了，因此最终触发了很快就会被遗忘的之字形的绘制。 “解决问题”甚至是讨人喜欢的表达“调试”可能都没有抓住重点：通过融合两组不同且同样重要的实践，它们可能无法充分捕捉程序员在定义适当的指令列表时经历的微妙的实际节奏。

Let the reader forgive me if I rave a little at this point of the chapter, but both technical and scientific practices as documented in these two sections provide such a refreshing perspective on computer programming that it is difficult for me to remain placid. We see indeed how the standard cognitive-behavioral framing of computer programming as a problem-solving process (cf. chapter 3) can be misleading. Programmers may never solve any problem; when confronted to a remote entity that refuses to generate electric pulses on data, they more or less collectively constitute a chain of reference that, if equipped enough, may indicate a problematic location, a location that, in turn, may trigger the enrollment of new actants and technical work-arounds of impasses. Nothing is solved; something is located, thus eventually triggering the drawing of a zigzag that will soon be forgotten. “Problem solving” and even the likable expression “debugging” may both miss the point: by amalgamating two different and equally important sets of practices, they may not adequately catch the subtle practical tempos a programmer goes through when defining appropriate lists of instructions.

然而，尽管我充满热情，但这个尝试性的模型仍然缺少一些关键的东西。事实上，这种“适当性”从何而来？它不是我偷偷从外部调用的东西，而没有定义它的属性吗？目前，它肯定是。幸运的是，这正是本章下一节的主题。

Yet, despite my enthusiasm, this tentative model still lacks something crucial. Indeed, where does this “appropriateness” come from? Is it not something I surreptitiously invoke from outside, without defining its attributes? At this point, it surely is. Fortunately, this is precisely the topic of the next section of this chapter.

附有场景

Attached to a Scenario

到目前为止，我们已经看到，编程可以被视为两组密切相关的实践的表达。第一组意味着铭文的乘法和对齐，以组装参考链可以提供有关远程实体的信息，这些实体的轨迹受到了不良影响。这些实践在某种程度上与构建科学事实所需的一些实验室实践相呼应。第二组实践——更难捕捉——意味着纳入新的行动者以绕过僵局。这些包含、授权和绕过的实践在某种程度上与运行技术项目所需的实践相呼应。从这一点来看，我们可以推测，在计算机编程过程中，科学和技术实践是紧密结合的，程序员不断地从一种模式转换到另一种模式。这种对计算机编程的尝试性但经验性的审视揭示了许多关键元素——铭文、引用链、僵局、弯路——大多数标准的计算机编程观点都没有强调这些元素。

We have seen so far that programming can be viewed as the expression of two sets of intimately related practices. The first set implies the multiplication and alignment of inscriptions in order to assemble chains of reference that can provide information about remote entities whose trajectories are affected in undesirable ways. These practices echo well, to some degree, with some of the laboratory practices required for the construction of scientific facts. The second set of practices—much more difficult to capture—implies the inclusion of new actants in order to get around impasses. These practices of inclusion, delegation, and bypassing echo well, to some degree, with practices required for the running of technical projects. From this point, we may conjecture that during a computer programming episode, scientific and technical practices are intimately articulated, the programmer constantly shifting from one mode to the other. This tentative but empirical look at computer programming unfolds many crucial elements—inscriptions, chains of reference, impasses, detours—that most standard takes on computer programming do not stress.

不过，在本章的这一点上，计算机编程的一些基本要素仍被视为理所当然。当我继续谈论“编程情节”时，是什么定义了这些情节的限制和范围？这些确定编程情节边界的“元指令”来自哪里？是什么风推动着程序员，让他们查询远程实体，注册执行者，并绕过僵局？在本章的前一节中，读者可能已经注意到，我偷偷使用了“编程项目”一词来谈论 DF 为 PROG 的组成部署的技术技能。但这种投射从何而来？在这一点上，这种愿望、这种渴望不应再被忽视。现在是时候解决投射和依附的问题了，没有这些问题，就不会有编程实践。

At this point of the chapter though, something essential to computer programming is still taken for granted. While I keep on talking about “programming episodes,” what defines the limits and the scope of such episodes? Where do these “meta-instructions” that establish the boundaries of the programming episodes come from? What is this wind that pushes programmers in the back, making them inquire into remote entities, enroll actants, and get around impasses? In the previous section of the chapter, readers may have noticed that I surreptitiously used the term “programming project” to speak about the technical skills DF was deploying for the composition of PROG. But where does this projection come from? At this point, this aspiration, this desire shall not be ignored anymore. It is time now to address the issues of projection and attachment without which there would simply be no programming practices.

露西·萨奇曼彻底探讨了计划与情境行动之间的关系，或者用她的话来说，“预测未来行动的效用以及这些预测对它们未详尽指定的进一步活动范围的依赖”（Suchman 2007，19；重点为原文）。她最初反对 20 世纪 80 年代中期的人工智能专家，这些专家倾向于将计划与行动之间的关系视为确定性的——前者严格定义后者——她提出了另一种观点，认为计划是设定范围的资源，但没有指定达到范围所需的行动。为了阐明她的主张，她使用了独木舟的例子：

Lucy Suchman thoroughly explored this relationship between projects and situated actions or, as she put it, “the utility of projecting future actions and the reliance of those projections on a further horizon of activity they do not exhaustively specify” (Suchman 2007, 19; emphasis in the original). Initially struggling against mid-1980s artificial intelligence experts who tended to consider the relationship between plans and actions as deterministic—the former rigorously defining the latter—she proposed an alternative view of plans as resources that set up horizons without specifying the actions required to reach them. To clarify her proposition, she used the example of canoe:

在计划乘独木舟穿越一系列急流时，人们很可能会坐在瀑布上方一段时间，计划下山。计划可能像这样：“我尽可能地向左走，试着绕过下一个急流。”这样的计划可能需要大量的考虑、讨论、模拟和重建。但无论多么详细，计划都无法真正让你的独木舟穿越瀑布。当真正涉及到应对水流和操纵独木舟的细节时，你实际上会放弃计划，并依靠你所能掌握的任何具体技能。在这种情况下，计划的目的不是让你的独木舟穿越急流，而是让你以某种方式定位自己，以便你能够获得最佳位置来使用那些具体技能，而归根结底，你的成功取决于这些具体技能。（Suchman 2007，72）

In planning to run a series of rapids in a canoe, one is very likely to sit for a while above the falls and plan one’s descent. The plan might go something like “I get as far over to the left as possible, try to make it around that next bunch.” A great deal of deliberation, discussion, simulation, and reconstruction may go into such a plan. But however detailed, the plan stops short of the actual business of getting your canoe through the falls. When it really comes down to the details of responding to currents and handling a canoe, you effectively abandon the plan and fall back on whatever embodied skills are available to you. The purpose of the plan in this case is not to get your canoe through the rapids, but rather to orient you in such a way that you can obtain the best possible position from which to use those embodied skills on which, in the final analysis, your success depends. (Suchman 2007, 72)

计划并不能决定行动。相反，通过暗示未来的方向，计划有助于在可感知的条件下表达技能。此外，基于萨奇曼的例子，我们还可以假设计划创造了另一个世界、另一个存在层面：通过讲述故事，计划表达了没有它们就无法存在的形象。在奔向急流之前，当我在瀑布上方表达我的计划时，我被投射到另一个空间（“进入急流”、“尽可能向左”）、另一个时间（“稍后”），以及其他人类和非人类行为者（“我，活着，在急流的尽头”、“独木舟，挣扎着绕过下一个急流”、“我希望能够逃离的湍急急流”）。在这方面，通过建立三重向外的转移（Latour 2013，234–257），进入其他空间和时间，并向其他行为者转移，计划也是帮助我们参与理想过程的叙述。

Plans do not determine actions. Rather, by suggesting future orientations, plans help express skills in appreciable conditions. Moreover, building on Suchman’s example, we can also assume that plans create something like another world, another layer of existence: by telling stories, plans express figures that could not exist without them. Before running the rapids, when I am expressing my plan above the fall, I am projected into another space (“into the rapids,” “as far over to the left as possible”), another time (“later”), and toward other human and nonhuman actants (“me, alive, at the end of the rapids,” “the canoe, struggling to get around the next bunch,” “the powerful rapids I—hopefully—managed to run”). In this respect, by establishing a triple shifting out (Latour 2013, 234–257) into other space and time, and toward other actants, plans are also narratives that help us engage into desirable processes.

然而，将计划定义为建立理想视野而不具体说明如何实现这些视野的叙述，这种定义仍然相当宽泛。这些叙述在哪些方面不同于儿童睡前故事或好莱坞大制作？计划叙述会带来哪些具体转变？我们该如何应对它们所建议的修改？为了更好地理解这些叙述（或者，我很快会称之为这些场景）的特殊性，我们将考虑为完成 PROG 而构建的叙述 DF。一个出发点可能是我们在上几节中关注的编程事件发生前两天。当时，我正在努力处理之前从众包任务中收集的数据。由于无法理解这些数据，我向实验室主任 (DIR) 寻求一些建议：

Yet this definition of plans as narratives establishing desirable horizons without specifying how to reach them is still quite loose. In what sense are these narratives different from, say, bedtime stories for children or Hollywood mega-productions? What specific transformations do plans-narratives institute? How do we address the modifications they suggest? To better understand the specificity of these narratives—or, as I will soon call them, these scenarios—we shall consider the narrative DF constructed for the completion of PROG. One point of departure could be two days before the programming episode we have followed in the last sections. At that time, I was struggling with the data I had previously collected from a crowdsourcing task. Unable to make sense of these data, I asked the director of the Lab (DIR) for some advice:

2016 年 2 月 4 日星期四，DIR 办公室

Thursday February 4, 2016, at the office of DIR

缩略词：  问题是，我仍在努力寻找能够理解工人绘制的矩形变化的措施，这些变化取决于图像。11^因为此时，我得到了这样的结果：

FJ:  The thing is that I am still struggling to find measures that could make sense of the variations of the rectangles drawn by the workers [and] depending on the images.¹¹ Because at this point, I have this kind of result:

[FJ 向 DIR 展示了他笔记本电脑上的图像，见图4.33 ]

[FJ shows images on his laptop to DIR, see figure 4.33]

缩略词：  但是矩形在大小和对齐方面都有所不同。也就是说，有些矩形对齐得很好，而且与图像相比较小；有些矩形对齐了，但尺寸不同；有些矩形对齐了，但分成了不同大小的组；还有一些矩形则分散在各处。

FJ:  But the rectangles vary both in terms of size and alignment. That is, some rectangles are well aligned and small compared to the image; others are aligned but vary in terms of dimensions; others are aligned but in groups of different sizes; and others are just spread out everywhere.

总监：  嗯，肯定有办法测量重叠程度。但无论如何，你应该能看到除这些之外的其他视图。你在这里看不到任何东西。

DIR:  Well, there’s surely a way to measure how much overlap there is. But in any case, you should get other views than these. You can’t see anything here.

…

…

方法有很多，但例如，您可以查看每个像素，看看它们在矩形中出现的频率。一旦您获得这些图表，我们就可以帮您找到解释这些变化的度量。

There are many ways; but for example, you could go through each pixel and see how often they are in a rectangle. And once you get these graphs, we can help you find a measure that explains the variations.

缩略词：  您的意思是，获取每个像素所属的矩形数量的相对差异？

FJ:  You mean, something like getting for each pixel, the relative difference of the number of rectangles they are part of?

总监：  是的。或者更确切地说，我猜在你的情况下，对于每幅图像，像素中属于一个矩形、两个矩形等的比例。……然后你可以得到灰度图像，或者像直方图这样的图表。例如，假设你给每个没有人标记的像素分配零，给每个只有一个工人标记的像素分配一，等等。你把这些加起来，你会得到一个最大值，或者像二十。然后你可以在零和一之间进行标准化，或者做其他事情。但至少现在，你应该从这些图像中获得更好的矩阵。

DIR:  Yes. Or rather, I guess in your case, for each image, the proportion of pixels that are part of one rectangle, two rectangles, and so on. … And then you can get gray-scale images, or graphs like histograms. For example, assume you’re giving zero to every pixel that is labeled by no one, one for every pixel that is labeled by only one worker, etc. You add this up and you’ll get a maximum or, like twenty. Then you can normalize between zero and one or do other things. But for now at least, you should get better matrices from these images.

DIR 的建议很明确：如果我想找到图像像素值与工人绘制的矩形之间的相关性，第一步就是通过设计更好的矩阵来简化收集的结果。但这些矩阵应该如何设计呢？这个问题是 PROG存在的理由：为了定义更简单/更好的矩阵，其值可以用图形表示，PROG 应该指示我的计算机转换每个图像及其相关矩形的值。简而言之，可以帮助我解释矩形分散/对齐的图形需要矩阵，而这些矩阵仍然需要由受过指导的计算机通过计算设计计算机。第一个进一步支持 PROG 制定的叙述或计划可以总结如下：“FJ 将让计算机组装矩阵，其像素值对应于每个像素所属的矩形数量。”

DIR’s advice was clear: if I wanted to find correlations between the pixel-values of the images and the rectangles drawn by the workers, the very first step was to simplify the collected results through the design of better matrices. But how should these matrices be designed? This issue was the raison d’être of PROG: in order to define simpler/better matrices whose values can be expressed by graphs, PROG should instruct my computer to transform the values of each image and its associated rectangles. In short, the graphs that could help me explain the dispersion/alignment of rectangles required matrices that still needed to be designed computationally by an instructed computer. The first narrative—or plan—that further supported the formulation of PROG can thus be summarized as such: “FJ shall make a computer assemble matrices whose pixel-values correspond to the number of rectangles each pixel is part of.”

我很快就尝试编写这个程序，以便更好地掌握我收集的数据，但很快我就发现自己无法用 Matlab 指定问题。第一步应该是什么？第二步又应该是什么？利用项目的帮助条款，我可以随时寻求帮助（参见上文），我给 DF 发送了一封电子邮件：

I soon tried to write this program that could help me have a better grip on the data I had collected but was soon confronted to my incapacity to specify the problem with Matlab. What should be the first step? And the second step? Using the project’s helping clause that allowed me to ask for help whenever I needed to (cf. above), I sent an email to DF:

2016 年 1 月 15 日，星期一。FJ 发给 DF 的电子邮件，标题为“与 Matlab 斗争…… ”

Monday, January 15, 2016. Email from FJ to DF, header “Struggling with Matlab. …”

你好，DF，

Hi DF,

对于我的项目，我需要单独处理每幅图像的每个像素，以便计算每个像素有多少个矩形。我想我有这个想法，但仍在努力用 Matlab 编写脚本。你能抽出时间帮我做一下吗？那太好了！

For my project I need to process each pixel of each image individually in order to count how many rectangles belong to each pixel. I got the idea, I think, but am still struggling with Matlab to write the script. Would you have some time to help me do it? That’d be great!

祝你有美好的一天，

Have a great day,

缩略词

FJ

2016 年 1 月 15 日，星期一。DF 发给 FJ 的电子邮件，标题为“与 Matlab 斗争…… ”

Monday, January 15, 2016. Email from DF to FJ, header “Struggling with Matlab. …”

你好，Florian，

Hi Florian,

没问题。那么今天下午呢？应该很容易。我们一起检查一下。

No problem. What about this afternoon then? It should be quite easy. We’ll check this together.

自由度

DF

2016 年 1 月 15 日，星期一。DF 发给 FJ 的电子邮件，标题为“与 Matlab 斗争…… ”

Monday, January, 15, 2016. Email from DF to FJ, header “Struggling with Matlab. …”

今天下午很好。我会在办公室。你想什么时候来都可以。

This afternoon is great. I’ll be in my office. Come whenever you want.

回头见！

See you then!

缩略词

FJ

几个小时后，DF 来到我的办公室。在开始编程之前，他告诉我他打算做什么：

A couple of hours later, DF arrived at my office. Before starting to program, he told me what he intended to do:

国防军：  “好吧，我想我知道如何计算。这应该不难。所以对于每个矩形，我们都有x和y坐标，对吗？”

DF:  “Well, I think I know how to compute this. It shouldn’t be difficult. So for each rectangle, we have the x and y coordinates right?”

缩略词：  “嗯，矩形由四个值定义”

FJ:  “Well, a rectangle is defined by four values”

国防军：  “是的，那么x和y [坐标] 然后是尺寸，对吗？”

DF:  “Yes so x and y [coordinates] and then the size, right?”

缩略词：  “是的。”

FJ:  “Yes.”

国防军：  “所以基本上我们有这个。”

DF:  “So basically we have this.”

[DF 开始在 FJ 的日志中书写]

[DF starts to write in FJ’s logbook]

国防军：  “还有这个，还有尺寸。所有这些定义了矩形。”

DF:  “And this, and then size. And all this defines the rectangle.”

[DF绘制图4.34 (A)]

[DF draws figure 4.34 (A)]

国防军：  “在这里 [指着图 4.34 (A)]，你用值 0 初始化矩阵的所有像素。然后迭代所有矩形。因此，对于图像的第一个矩形 [开始在 FJ 的日志中绘制]，你有坐标，并且你检查矩阵的哪些像素位于矩形中。”

DF:  “Here [pointing at figure 4.34 (A)], you initialize all pixels of the matrix with the value 0. Then you iterate on all rectangles. So for the first rectangle of the image [starts to draw in FJ’s logbook], you have the coordinates and you check what pixels of the matrix are in the rectangle.”

[DF绘制图4.34 (B1)]

[DF draws figure 4.34 (B1)]

国防军：  “然后你为矩阵中的这些像素添加一个。然后你对可能在这里的第二个矩形执行相同的操作 [开始在 FJ 的日志中绘制]。”

DF:  “And you add one for these pixels in the matrix. And then you do the same for the second rectangle [starts to draw in FJ’s logbook] that might be here.”

[DF绘制图4.34 (B2)]

[DF draws figure 4.34 (B2)]

国防军：  “你还要为所有这些像素加一。所以这里 [指着图 4.34 (B2)]，矩阵中的一些像素的值为 0，一些像素的值为 1，还有一些像素的值为 2。”

DF:  “And you also add one for all these pixels. So here [pointing at figure 4.34 (B2)], some pixels in the matrix will have the value 0, some will have the value 1, and some others will have the value 2.”

缩略词：  “好的我明白了。”

FJ:  “OK, I see.”

国防军：  “然后对所有矩形都执行此操作。一旦你有了一个适用于一张图片的脚本，就很容易调整它（脚本）来处理所有图片。”

DF:  “And you do this for all the rectangles. And once you have a script that works for one image, it’s easy to adapt it [the script] to go through all the images.”

缩略词：  “当然。”

FJ:  “Sure.”

国防军：  “当你拥有这些值为 0、1、2 等的矩阵时，你就可以制作所有你想要的图形，比如灰度图像或直方图 [在 FJ 的日志中绘制] 就像这样。”

DF:  “And well, when you have these matrices with values 0, 1, 2, etc., you can make all the graphs you want like gray-scale images or histograms [draws in FJ’s logbook] like this.”

[DF绘制图4.34 (C)]

[DF draws figure 4.34 (C)]

国防军：  “其中x是矩形的数量，y 是像素的数量。”

DF:  “Where x is the number of rectangles and y the number of pixels.”

此时，PROG 的叙述变得更加丰富。从“FJ 应让计算机创建矩阵，其像素值对应于它们所属的矩形数量”，变成了“对于每一幅图像，DF 应首先让计算机使用图像的维度创建一个空矩阵，然后根据相关 .txt 文件中定义的坐标定义该图像的第一个矩形，然后将此矩形添加到矩阵中，然后定义第二个矩形，然后将其添加到矩阵中，对图像的每个矩形依此类推。”虽然这个主题与 Suchman (2007, 72) 的独木舟例子略有不同，但 DF 的叙述也可以作为一种资源，它设置了一个地平线，但没有指定到达地平线所需的操作。没有提到如何定义空矩阵、如何定义矩形以及如何使用这些矩形增加矩阵。然而，总而言之，这些步骤的堆积建立了后续操作应该尝试达到的理想未来。此外，与 Suchman 的例子类似，DF 的叙述也创建了另一层存在。他的故事将我们带入另一个时间（“几分钟后”）、另一个空间（“在 Matlab IDE 前面”）和其他行动者（“增量矩阵”、“灰度图像”、“直方图”、“FJ 能够借助新程序生成有意义的图表”）。

At this point, the narrative of PROG has thickened. From “FJ shall make a computer create matrices whose pixel values correspond to the number of rectangles they are part of,” it has become “for every image, DF shall first make a computer use the dimension of the image to create an empty matrix, then define the first rectangle of this image according to its coordinates as defined in its correlated .txt file, then add this rectangle to the matrix, then define the second rectangle, then add it to the matrix, and so on for every rectangle of the image.” Even though the topic is slightly different from Suchman’s (2007, 72) example of canoe, DF’s narrative also works as a resource that sets up a horizon without specifying the actions required to reach it. Nothing is said about how to define the empty matrix, how to define a rectangle, and how to increment the matrix with these rectangles. Yet, altogether, the pileup of these steps institutes a desired future that the following actions should try to reach. Moreover, similar to Suchman’s example, DF’s narrative also creates another layer of existence. His story projects us into another time (“in a couple of minutes”), another space (“in front of the Matlab IDE”), and toward other actants (“incremented matrices,” “gray-scale images,” “histograms,” “FJ being able to produce meaningful graphs thanks to the new program”).

但是，结合本章最后两节的论述，DF 的叙事也表明，设定期望未来的叙事与儿童睡前故事或好莱坞大片之间存在重要区别。在叙事表达完毕之后（即，在叙事被投射到其他时间、其他地点和其他行动者身上之后），希望孩子们入睡，观众离开电影院继续他们的工作，此时 DF 的叙事仍然对他有控制力。DF 的叙事不仅仅是建立了向其他时空和其他行动者的三重转移，还吸引着DF；它要求 DF 做事。从这个意义上说，一旦 DF 表达了叙事，他就会发现自己同时处于两个位置：他是叙事的作者，可以随时修改叙事；但他也是演员，必须遵循他刚刚表达的叙事（Latour 2013, 391）。根据 Austin (1975) 和 STS 的最新作品 (Barad 2007)，我们可以将这些叙事视为表演性的，因为它们会吸引那些表达它们的人。在我们的例子中，DF 掌握了叙事，但也被叙事所控制。

But DF’s narrative—when considered in the light of the last two sections of this chapter—also suggests an important difference between narratives that institute desired futures and, say, bedtime stories for children or Hollywood mega-productions. When after the narrative has been expressed—that is, after having been projected into other times, other locations, and toward other actants—hopefully children fall into sleep and spectators leave the movie theater to carry on their occupations, DF’s narrative still has a hold on him. More than just establishing a triple shifting out into other space and time and toward other actants, DF’s narrative engages DF; it asks DF to do things. In this sense, as soon as DF expresses the narrative, he finds himself simultaneously in two positions: he is the writer of the narrative who can modify it any time he wants but also the actor who has to follow the narrative he has just expressed (Latour 2013, 391). Following Austin (1975) and recent works in STS (Barad 2007), we can consider these narratives as performative in the sense that they engage those who articulate them. In our case, DF holds the narrative but is also held by it.

为了强调这些特定叙事的文学和表演维度，即它们对于计算机编程至关重要，因为它们设定了期望的视野，从而支持铭文和技术迂回，我将它们称为^场景。12电影摄影的内涵是自愿的。事实上，场景——在电影或计算机编程的情况下——是一种叙事：它讲述了一个故事，因此实例化了开头、结尾、情节和人物，所有这些都具有本体论的权重。其次，在这两种情况下，场景都是表演性的：它既对被要求将其转化为电影的电影导演有影响，也对试图将其变成实际计算机程序的程序员有影响。第三，如果一个场景粗略地描述了电影的连续场景或计算机程序的连续步骤，它几乎没有提到如何拍摄这些场景或实施这些步骤。虽然在这两种情况下，场景都描绘了理想的视野，但为了达到这些视野，几乎还需要做所有事情。第四，如果情节、步骤、人物或变量是由剧本描述的，那么没有什么可以阻止电影导演、程序员、电影明星或顽固的指令修改其部分构成要素。无论是在计算机编程还是电影制作中，都可以重新审视一个场景，以更好地考虑不可预测的意外情况。

To underline the literary and performative dimensions of these particular narratives that are crucial for computer programming—since they institute a desired horizon to be achieved, hence supporting both alignments of inscriptions and technical detours—I shall call them scenarios.¹² The cinematographic connotation is voluntary. Indeed, a scenario—in the case of cinema or computer programming—is a narrative: it tells a story and therefore instantiates a beginning, an end, a plot, and characters that all possess ontological weights. Second, in both cases, a scenario is performative: it has a hold on both the movie director who is asked to transform it into a movie as well as on the programmer who tries to make it become an actual computer program. Third, if a scenario roughly describes the successive scenes of a movie or the successive steps of a computer program, it says almost nothing about how to shoot these scenes or implement these steps. While in both cases, the scenario draws desirable horizons, almost everything still needs to be done in order to reach them. Fourth, if the plot, steps, characters, or variables are described by the scenario, nothing prevents the movie director, the programmer, movie stars, or recalcitrant instructions to modify some of its constitutive elements. In both computer programming and movie production, a scenario can be revisited to better take into account unpredictable contingencies.

虽然场景不足以组装计算机程序，但对于计算机编程来说，场景仍然至关重要。这些灵活而又富有表现力的叙事资源为程序员建立了视野，同时也被它们所束缚，从而确立了计算机编程情节的界限。场景既触发了铭文和技术弯路的排列，又与之融合在一起；总之，它们形成了我们现在可以考虑其所有曲折的编程过程。

While they are not sufficient to assemble computer programs, scenarios are nonetheless crucial for computer programming. These flexible yet performative narrative resources institute horizons on which programmers can hold—while being held by them—thus establishing the boundaries of computer programming episodes. Scenarios both trigger and are blended with alignments of inscriptions and technical detours; altogether, they form programming courses of action we can now consider in all their sinuosity.

但同样，在这一点上，仍然缺少一些东西。我们非常接近但还没有到达那里。如果“场景”的概念有助于更好地理解是什么帮助 DF 在科学和技术实践模式之间转变，从而构建我们之前遵循的编程序列，那么它并不能让我们理解为什么DF 想要参与其中。如果场景提供了编程情节的框架和能量，那么这种能量最初来自哪里？

But again, at this point, something is still missing. We are very close but are not there yet. If the notion of “scenario” is useful to better understand what helped DF shift between scientific and technical modes of practice, thus framing the programming sequences we have previously followed, it does not make us understand why DF wanted to engage himself in it. If scenarios provide the frame and the energy of programming episodes, where does this energy initially come from?

某种东西肯定充斥着场景，使它们“运转起来”，产生或多或少令人愉悦的影响：我们又该如何看待它们呢？如果场景提供了视野，它们本身并不允许抓住编程情节中产生的东西。INT 的固执、行为者的多重包含以及僵局的无数解决方法；所有这些——在行动的中间——都是非常不确定的。但是，当程序实现了情节开始时所希望的——或在情节过程中进行了修改——就会发生一些事情，而这些事情不能归结为允许它发生的后果。这是依恋社会学对品味社会科学的重要贡献：将心爱的物品归结为欣赏的条件——社会或物质——并不能告诉我们任何关于物品本身的信息（Hennion 2015, 2017）。虽然一个物体——一幅画、一段音乐、一个计算机程序——是被构造的，但它本身也存在。或者甚至更多；随着它的构造，它的存在更加强烈。但我们如何才能抓住对构成对象的这种欣赏呢？在我们的案例中，我们如何看待PROG的兴起？我们或许可以参考DF在编程片段结尾告诉我的话：

Something is definitely overflowing scenarios, making them “put into gear” more or less delightful affects: how do we consider them as well? If scenarios give horizons, they do not by themselves allow to grab what arises from programming episodes. INT’s stubbornness, the multiple inclusions of actants, and the numerous work-arounds of impasses; all of this—in the middle of the action—is terribly uncertain. But when the program accomplishes what was hoped for at the beginning of the episode—or modified during the episode—something is happening that cannot be reduced to the consequence of what allowed it to happen. This is the important contribution of the sociology of attachments against the social science of taste: reducing beloved objects to the conditions—social or material—of their appreciations tells us nothing about the objects themselves (Hennion 2015, 2017). While an object—a painting, a piece of music, a computer program—is constructed, it also exists in its own right. Or perhaps even more; as it is constructed, it exists more intensely. But how do we grab this appreciation of the constituted object? In our case, how do we consider the upsurging of PROG? We may perhaps refer to what DF tells me at the end of the programming episode:

缩略词：  “好吧，谢谢。你的耐心总是让我印象深刻。”

FJ:  “Well, thanks. I’m always impressed by your patience.”

国防军：  “不用客气。速度很快。你知道，我很喜欢它，所以这不是问题。”

DF:  “You’re welcome. It was quick. And you know, I love it so it’s not a problem.”

缩略词：  “你喜欢花时间在这些代码上吗？”

FJ:  “You love spending time on these lines of code?”

国防军：  “当然。这很有趣。我真正喜欢的是，你永远不应该失去线索。当剧本完成任务时，就意味着你没有失去线索。”

DF:  “Sure. It’s fun. What I really like is that you should never lose the thread. And when the script does the thing, it means you didn’t lose it.”

这段摘录能告诉我们计算机编程的影响吗？场景的概念本身似乎无法让我们更清楚地理解 PROG 在组装后对 DF 的作用。但是，遵循 DF 并使用场景作为垫脚石，它有助于让一些可爱的事情出现：能够不断评估已经完成的事情与仍需完成的事情。这是 DF 始终需要抓住的，他试图永不失去的线索：这个场景暗示了一条道路、一个情节，但也没有说明如何遵循它。通过追踪自己的道路来追随一个故事：通过到达某物来建立某物的一种奇妙体验。但这种到达，这种进入地平线的途径——人们不应该简单地将其视为实现先前预测的事物的满足感。认真对待 DF——以及参加其他“帮助会议”的其他实验室合作者——我们可以将其视为不断评估的渐近线。 “这个”必须先做，然后是“这个”，然后是“这个”，现在，当然，在下一个影响场景出现之前，没有其他事情可做了。计算机编程影响的特殊性可能在于这种暂时的“没有其他事情”反复出现。

What may this excerpt tell us about the affects of computer programming? The notion of scenario seems, by itself, unable to provide a clearer understanding of what PROG, once assembled, does to DF. But, following DF and using the scenario as stepping stone, it helps to make appear something lovable: being able to constantly evaluate what has been done against what still needs to be done. This is what DF steadily needs to grab, the thread he tries never to lose: this scenario suggests a path, a plot, but also says nothing about how to follow it. Following a story by tracing his own path: a curious experience of establishing something by reaching it. But this reach, this access to the horizon—one should not simply consider it as the satisfaction of realizing something that was previously projected. Taking DF seriously—but also other Lab collaborators who participated in other “helping sessions”—we may consider it as the asymptote of a constant evaluation. “This” had to be done, then “this,” then “this,” and now, there is nothing else to do until the next affect-bearing scenario, of course. The specificity of the affects of computer programming may lie in the recurrent upsurging of this temporary “nothing else.”

这只是一个关于将程序员与他们可能安装的脚本绑定在一起的依恋关系的大胆提议（ Latour 2013，151-178；Souriau [1943] 2015）。显然需要进行更系统的研究来丰富上述推测。但请读者再次不要忘记，本章的目标之一，除了分析野心之外，还在于指出计算机编程情境实践研究的创新途径。从这个意义上讲，研究场景的形成及其与它们可能暗示（但严格来说并非产生）的依恋关系的复杂关系，可能是探究是什么促使程序员行动的一种相关方式，有时甚至会花费大量无偿（或绕道而行）的时间在不确定的免费和开源软件项目上。鉴于程序员对场景的执着，Demazière 、 Horn 和 Zune (2007, 35) 所说的“自由软件开发之谜”——从短暂的参与中产生连贯的编程结果的能力——可以用另一种方式来解决。虽然这些自愿集体之间纠缠不清的监管模式对于自由和开源软件的实际生产当然很重要，但这些安排也应该从它们所激发的激情的角度加以考虑。当一个场景通过计算机脚本实现时，究竟发生了什么？这种情感事件是否只能归结为使其成为可能的组织过程（Demazière 、 Horn 和 Zune 2007）、个人激励（Lerner 和 Tirole 2002）或意识形态（Elliott 和 Scacchi 2008）？DF 的情感火花中是否也有一些东西可以促进程序员社区的形成和维护？编程工作的整个生态——无论是免费的、开源的还是企业的——可能也值得从程序员在编写编号的指令列表时所追求的东西的角度加以考虑。

This is only an adventurous proposition about the attachments that bind programmers to the scripts they may instaure (Latour 2013, 151–178; Souriau [1943] 2015). More systematic studies are obviously necessary to enrich the above speculations. But let the reader not forget, once again, that one goal of this chapter, besides its analytical ambitions, is also to point to innovative avenues of research on computer programming situated practices. In that sense, looking at the formation of scenarios and their complex relationships with the attachments they may suggest—but not strictly produce—could be a relevant way to inquire into what moves programmers, sometimes to the point of spending huge amount of unpaid (or detoured) hours on uncertain free and open-source software projects. In the light of programmers’ attachments to scenarios, what Demazière, Horn, and Zune (2007, 35) called the “enigma of free software development”—the ability to produce coherent programming results from evanescent involvement—could, for example, be tackled in an alternative way. While entangled modes of regulations among these voluntary collectives are certainly important for the actual production of free and open-source software, these arrangements may also benefit from being considered in the light of the passions they make exist. What is indeed happening when a scenario is realized through a computer script? Can such an affective event only be reduced to the organizational processes (Demazière, Horn, and Zune 2007), individual incentives (Lerner and Tirole 2002), or ideologies (Elliott and Scacchi 2008) that made it possible? Is there not something in DF’s emotive spark that may also contribute to the formation and maintenance of programmers’ communities? It is the whole ecology of programming work—be it free, open-source, or corporate—that may deserve to be considered also in the light of what programmers are after when they are writing numbered lists of instructions.

尽管篇幅冗长而曲折，但我想在第二部分中提出的观点非常简单。一旦我们研究计算机编程的行动过程，我们就会发现它们涉及铭文的对齐、僵局的解决方法以及场景的定义。这三种实践模式密切相关：解决僵局意味着对问题现象进行定位，而这种现象本身需要一种场景才能被视为问题。DF 和更普遍地说，程序员不断地从一种模式切换到另一种模式，直到暂时实现他们想要的叙述。

Despite its lengthy and tortuous aspect, the point I wanted to make in this part II is quite simple. Once we inquire into computer programming courses of action, we see that they engage the alignment of inscriptions, the work-around of impasses, and the definition of scenarios. These three modes of practices are intimately related: Working around impasses implies the localization of a problematic phenomenon that itself requires a scenario to be considered problematic. DF and more generally, perhaps, programmers constantly shift from one mode to the other until temporally realizing their desired narratives.

主要困难在于区分编程过程和结果所需的准备工作。由于我们在第 3 章中介绍的复杂原因，人类认知和编程计算机之间逐渐形成了令人困惑的混合。这种混淆反过来又导致了重要的误解，例如编程的认知研究，这些研究最终变得同义反复，因为它们假设了它们试图解释的东西的存在。由于我想分析编程的情境实践，我不得不远离认知主义，接受非常简约但强大的行动主义，这种行动主义认为认知是我们掌握局部环境可供性的过程。

The main difficulty lay in the preparatory work required to distinguish the process of programming from its result. For complicated reasons we covered in chapter 3, a confusing mix has progressively been established between human cognition and programmed computers. This confusion led, in turn, to important misunderstandings such as cognitive studies of programming that ended up being tautological as they supposed the existence of what they tried to account for. As I wanted to analyze the situated practice of programming, I had to distance myself from cognitivism and embrace very minimal, yet powerful, enactivism that considers cognition as the process by which we grasp affordances of local environments.

不幸的是，我只能在计算机编程实践的边缘上玩，许多问题都没有得到解答。关于铭文的对齐，了解更多关于参与程序员铭文乘法和表达的不同模式、组织甚至机构的信息将大有裨益。关于解决僵局，更彻底地探索支持识别和登记新行为者的设备怎么样？这甚至可能导致创新的编程设备和设备。关于场景，我很快就会记录一些特定的、易于转换的场景的形成。但鉴于计算机编程的魅力及其对当代社会的重要性，我希望有更多的研究记录有时让编程的乐趣浮现出来的行动。在这个对算法（似乎依赖于地面实况和编程活动的实体）充满争议的时代，我相信这些是至关重要的研究方向。

Unfortunately, I could play only at the edge of computer programming practices, and many questions were left unanswered. Regarding the alignment of inscriptions, it would be insightful to learn more about the different modalities, organizations, and even institutions that participate in a programmer’s multiplications and articulations of inscriptions. Regarding the working around of impasses, what about exploring more thoroughly the equipment that supports the identification and enrollment of new actants? This may even lead to innovative programming devices and equipment. Concerning scenarios, I will soon document the formation of some specific, easily transposable ones. But in light of the fascination exerted by computer programming as well as its importance for contemporary societies, I wish there were more studies documenting the actions that sometimes make the joy of programming emerge. In these times of controversies over algorithms—entities that seem to rely on ground-truthing and programming activities—these are, I believe, crucial research directions.

笔记

Notes

1. 1.为了进行这个项目，我必须精通 Python、PHP、JavaScript 和 Matlab 编程语言。

2. 1.  To conduct this project, I had to become competent in Python, PHP, JavaScript, and Matlab programming languages.

3. 2.值得注意的是，这种逐行翻译是程序员所经历的。在 INT 和大多数其他解释器的轨迹中，书面符号的编号列表被翻译成抽象语法树，而这并不总是保留编辑器的逐行表示。

4. 2.  It is important to note that this line-by-line translation is what is experienced by the programmer. In the trajectory of INT and most other interpreters, the numbered list of written symbols is translated into an abstract syntax tree that does not always conserve the line-by-line representation of the Editor.

5. 3.很难确切知道 INT 在 T1 时如何处理这三个值。它可能默认认为只有图像大小的前两个值（宽度和高度）通常很重要。

6. 3.  It is difficult to know exactly how INT managed to deal with these three values at T1. It may by default consider that only the first two values of image-size—width and height—generally matter.

7. 4.在 Matlab 编程语言中，每个非条件语句，且不以分号结尾的语句，默认情况下，解释器都会在命令窗口中打印出来。这与许多其他高级编程语言不同，这些语言的打印操作应由指令指定（通常是指令“打印”）。

8. 4.  In the Matlab programming language, every statement that is not conditional and that does not end with an semicolon is, by default, printed by the interpreter in the Command Window. This is different from many other high-level programming languages for which printing operations should be specified by an instruction (typically, the instruction “print”).

9. 5.在第5章中，我将探讨数学知识的形成，并更彻底地研究STS提出的科学事实的形成。

10. 5.  In chapter 5, where I will consider the formation of mathematical knowledge, I will more thoroughly examine the shaping of scientific facts as proposed by STS.

11. 6.这可能是软件研究的一个局限性，例如 Fuller (2008) 和《计算文化》杂志所提出的那样。通过考虑完整的代码，这些研究往往忽略了导致代码完成的实际操作。当然，这种观察仍然很重要，因为它使我们能够考虑与软件相关的文化产品的性能效果，这是我的行动导向方法无法完全做到的。

12. 6.  This may be a limitation of Software Studies, as for example presented in Fuller (2008) and in the journal Computational Culture. By considering completed code, these studies tend to overlook the practical operations that led to the completion of the code. Of course, this glance remains important as it allows us to consider the performative effects of software-related cultural products, something my action-oriented method is not quite able to do.

13. 7. Ullman (2012b) 以文学的方式详细记录了程序测试中组装引用链所需的连续操作。

14. 7.  The successive operations required to assemble chains of reference in the case of program-testing are well documented, though in a literary way, by Ullman (2012b).

15. 8 . 值得注意的是，DF 的对齐实践将因 Matlab 的下一版本而得到极大促进。事实上，2017 版 Matlab解释器在矩阵增量过程中自动识别这种类型的维度错误，并直接指出相关的断点，即发生问题的行（在我们的例子中是第 9 行）。

16. 8.  It is interesting to note that DF’s alignment practices would have been greatly facilitated by the next version of Matlab. Indeed, the 2017 version of Matlab’s interpreter automatically recognizes this type of dimension error during matrix incrementation processes and directly indicates the related breakpoint, the line at which the problem occurred (in our case, at line 9).

17. 9.唐纳德·克努斯 (Donald Knuth) 是最著名的编程理论家之一，他提出了“文学编程”的概念，强调了程序可理解性的重要性：这是一种计算机编程方法，主要侧重于向其他程序员解释程序，而不是“仅仅”指导计算机。

18. 9.  Donald Knuth, one of the most prominent programming theorists, stressed the importance of program intelligibility by proposing the notion of literate programming: a computer programming method that primarily focuses on the task of explaining programs to fellow programmers rather than “just” instructing computers.

19. 10.据我所知，只有三个例外：Vinck (1991)、Latour (2006) 和 Latour (2010b)。

20. 10.  To my knowledge, there are only three exceptions: Vinck (1991), Latour (2006), and Latour (2010b).

21. 11.本讨论根据 2015 年 11 月至 2016 年 3 月航海日志 8 中的记录重建。

22. 11.  This discussion has been reconstructed from notes in Logbook 8, November 2015–March 2016.

23. 12.一些 STS 作者使用术语“脚本”来定义这些特定的叙述，这些叙述会吸引那些阐述它们的人（Akrich 1989；Latour 2013）。如果我使用术语“场景”，主要是为了清晰起见，因为“脚本”经常被计算机科学家和程序员（包括我在本书中）用来描述诸如 PROG 之类的小程序。

24. 12.  Some STS authors use the term “script” to define these particular narratives that engage those who enunciate them (Akrich 1989; Latour 2013). If I use the term “scenario,” it is mainly for sake of clarity as “script” is often used by computer scientists and programmers—and myself in this book—to describe small programs such as PROG.

数学在哪里？

Where Is the Math?

如果我们想更好地理解数学实体（公式、定理、猜想、方程式）是如何被操纵的，以及它们与基本事实和编程语言之间的关系，我们首先需要更好地理解它们来自哪里。这些实体肯定不是独立存在的；它们需要由人们在特定的地方组装起来。这些地方在哪里？这些人是谁？他们做什么？

If we want to better understand how mathematical entities (formulas, theorems, conjectures, equations) are manipulated and related to ground truths and programming languages, we first need to better understand where they come from. Such entities surely do not exist by themselves; they need to be assembled by people in specific designated places. Where are these places? Who are these people, and what do they do?

这些琐碎的问题会引出许多不同的答案。这就是为什么处理数学可能会很危险的原因之一：我们应该从哪里开始？从古希腊的数学（Heath 1981a、1981b；Netz 2003）开始？从中世纪伊斯兰的数学（Berggren 1986；Netz 2004）开始？从巴洛克连续变化的数学（Bardi 2007；Boyer 1959）开始？但如果我们使用形容词“巴洛克”，我们已经以一种相当有针对性的方式定义了十七世纪（Deleuze 1992）。我们应该然后关注更现代的数学，如集合论 (Ferreir ós 2007; Tiles 2004)、魏尔斯特拉斯函数 (Bottazzini 1986) 以及随后在 20 世纪初震撼数学的“基础危机” (Ewald 2007; Ferreir ós 2008; Hesseling 2004; Mancosu 1997)？但我们所说的“数学”到底是什么意思？我们是指数学文本吗(Rotman 1995, 2006; Sha 2005)？我们是指莱布尼茨 (Antognazza 2011)、高斯 (Tent 2006) 或康托 (Dauben 1990) 等著名数学家吗？我们指的是那些试图定义数学是什么的数学哲学吗（Aspray and Kitcher 1988；Corfield 2006；Hacking 2014）？我们的头开始晕眩，头晕目眩。但还没有结束！事实上，我们谈论的是算术（Husserl 2012）、代数（Everest 2007）、几何（Netz 2003；Serres 1995、2002）还是逻辑（Fisher 2007；Rosental 2003）？也许我们谈论的是从数字到逻辑的演变（Kline 1990a）、从逻辑到几何（Kline 1990b；Netz 2003）、从几何到代数（Kline 1990c；Netz 2004）？甚至在算术、几何、代数或逻辑中，我们谈论的是定理（Villani 2016）、证明（Lakatos 1976；MacKenzie 1999、2004、2006）还是猜想（O'Shea 2008）？我们不知道。我们迷失在问题中，而这些问题的唯一表述让我们想做其他事情。但我们做不到；我们必须找到一种方法来解决数学问题，因为它对于算法的构成似乎很重要。我们该怎么做呢？

Such trivial questions lead to many, many heterogeneous answers. This is one reason why dealing with mathematics can be dangerous: Where shall we start? From the mathematics of ancient Greece (Heath 1981a, 1981b; Netz 2003)? From mathematics of medieval Islam (Berggren 1986; Netz 2004)? From baroque mathematics of continuous change (Bardi 2007; Boyer 1959)? But if we use the adjective “baroque,” we already define the seventeenth century in quite an orientated way (Deleuze 1992). Shall we then focus on more contemporary mathematics such as set theory (Ferreirós 2007; Tiles 2004), Weierstrass functions (Bottazzini 1986), and the subsequent “crisis of foundations” that shook up mathematics at the beginning of the twentieth century (Ewald 2007; Ferreirós 2008; Hesseling 2004; Mancosu 1997)? But what do we mean by “mathematics” anyway? Do we mean mathematical texts (Rotman 1995, 2006; Sha 2005)? Do we mean famous mathematicians such as Leibniz (Antognazza 2011), Gauss (Tent 2006), or Cantor (Dauben 1990)? Do we mean philosophies of mathematics that try to define what mathematics is (Aspray and Kitcher 1988; Corfield 2006; Hacking 2014)? Our head is spinning and we start to feel dizzy. But it is not over yet! Indeed, are we talking about arithmetic (Husserl 2012), algebra (Everest 2007), geometry (Netz 2003; Serres 1995, 2002), or logic (Fisher 2007; Rosental 2003)? Maybe are we talking about the evolution from numbers to logic (Kline 1990a), from logic to geometry (Kline 1990b; Netz 2003), from geometry to algebra (Kline 1990c; Netz 2004)? And even within arithmetic, geometry, algebra, or logic, are we talking about theorems (Villani 2016), proofs (Lakatos 1976; MacKenzie 1999, 2004, 2006) or conjectures (O’Shea 2008)? We do not know. We are lost in questions whose only enunciation makes us want to do something else. But we cannot; we must find a way to address mathematics as it seems important for the constitution of algorithms. How can we do so?

避免这种混乱的循环的一种方法是先从一些非常基本的假设开始。当然，我们必须发展这些假设，并用具体的例子来证明它们。要做到这一点，我们可能需要动员令我们害怕的庞大数学文献中的一小部分。一步一步，一个假设接一个假设——再加上一些 STS 假设——我们最终可能会得到一个数学知识的操作定义，足以完成我们的具体任务：解释计算机科学家在尝试组装新算法时有时能够动员他们的数学家同事先前形成的经过认证的命题的方式。我们当然不需要彻底改变我们对这些有时被称为“定理”、“猜想”或“公式”的有力陈述的理解。如果我们能塑造出数学家所做之事的一个简单版本（而不是数学是什么），我们最后的任务——解释制定实践——将得到极大的便利。

One way to avoid this spiral of confusion could be to start from some very basic hypotheses. We would, of course, have to develop these hypotheses and justify them by using concrete examples. To do this, we may need to mobilize a tiny part of the gigantic mathematics literature that scares us. One step after the other, one hypothesis after the other—coupled with some STS assumptions—we may end up with an operative definition of mathematical knowledge that could suffice to achieve our specific task: accounting for the way that computer scientists, when they try to assemble new algorithms, are sometimes able to mobilize certified propositions previously shaped by their mathematician colleagues. We surely do not need to revolutionize our understanding of these powerful statements we sometimes call “theorems,” “conjectures,” or “formulas.” If we just manage to shape one simple version of what mathematicians do (instead of what mathematics is), our last duty—accounting for formulating practices—will be greatly facilitated.

相对定罪强度的书面声明

Written Claims of Relative Conviction Strengths

为了启动我们的操作化练习并形成我们的第一个假设，让我们从三个围绕数学概念的场景开始：¹

To initiate our operationalization exercise and shape our first hypotheses, let us start with three scenes that all gravitate around mathematical notions:¹

场景 1

Scene 1

1994 年 1 月。查尔斯·埃尔坎 (Charles Elkan) 陷入了困境：他的定理表明模糊逻辑系统只能表达两个真值，这一定理引起了很大争议。2^哪里出了问题？他的定理在第十一届全国人工智能大会上的首次展示非常顺利。随后发表在会议论文集上的论文甚至被选为“最佳论文奖” (Elkan 1993)。程序委员会对证明的优雅性及其对专家系统进一步发展的意义表示赞赏。一切都已准备就绪，他的定理将被接受。但许多逻辑学家同事——他们没有参加会议，但阅读了麻省理工学院出版社出版的一些会议论文集——对此感到非常不满。埃尔坎甚至可以在新成立的互联网论坛“comp.ai.fuzzy”上表达他们的不满，该论坛致力于模糊逻辑理论和系统的高级讨论。批评很严厉。有人说——并试图证明——埃尔坎的基本假设是有缺陷的。其他人指责他故意削弱模糊逻辑，因为这是对古老、“陈旧”的古典逻辑的威胁。一些同事甚至怀疑他是头脑简单的亚里士多德主义者！正如他的一位朋友建议的那样，埃尔坎现在应该“冷静下来”，发表一个“更流畅”的定理版本，其中可能包括一些最中肯的批评。

January 1994. Charles Elkan is in turmoil: his theorem demonstrating that only two truth values can be expressed by a system of fuzzy logic is highly contested.² What went wrong? The initial presentation of his theorem at the Eleventh National Conference on Artificial Intelligence went very well. The paper that further appeared in the conference proceedings was even selected for the “Best Written Paper Award” (Elkan 1993). The program committee saluted the elegance of the proof as well as its significance for further developments in expert systems. Everything was in place for his theorem to be accepted. But many logician colleagues—who did not attend the conference but did read some of its proceedings published by MIT Press—are quite upset. Elkan can even follow their dissatisfaction on the newly established internet forum “comp.ai.fuzzy” that is dedicated to advanced discussions in fuzzy logic theories and systems. The critiques are harsh. Some say—and try to demonstrate—that Elkan’s basic hypotheses are flawed. Others accuse him of deliberately weakening fuzzy logic as it is a threat to old, “dusty” classical logic. Some colleagues even suspect him to be a thick-headed Aristotelian! As one of his friends advises him, Elkan should now “cool things down” and publish a “smoother” version of his theorem that could include some of its soundest critiques.

场景 2

Scene 2

1890 年夏天。阿尔弗雷德·肯普感到困惑；³虽然并不是因为珀西·希伍德最近设法在肯普之前发表在《美国数学杂志》（Heawood 1890；Kempe 1879）上的四色猜想的证明中找到了缺陷。希伍德做得很好，被驳斥也是游戏的一部分。不，更重要的是，尽管他的证明被证明是错误的，但肯普并不认为弗朗西斯·格思里 1852 年的坦率命题——即四种颜色足以给任何在平面上绘制的地图着色，使得没有相邻的假设许多国家的颜色相同——这是错的。但是，如此基本的直觉怎么会导致如此大的困难呢？难道数学家没有工具来证明这个猜想并一劳永逸地使其成为定理吗？“可怜的希伍德，”肯普想。“他现在沉迷于此，就像我十五年前一样。他最好放弃它；四色理论已经过时了。”

Summer of 1890. Alfred Kempe is puzzled;³ although not really because Percy Heawood recently managed to find a flaw in the proof of the four colors conjecture Kempe previously published in the American Journal of Mathematics (Heawood 1890; Kempe 1879). Heawood did a great job, and being refuted is part of the game anyway. No, it is more that even though his proof was shown to be erroneous, Kempe does not think that Francis Guthrie’s 1852 candid proposition—that says that four colors suffice to color any map drawn on a plane in such a way that no neighboring countries have the same color—is wrong. But how could such a basic intuition lead to such great difficulties? Do mathematicians not have the tools to prove this conjecture and make it a theorem once and for all? “Poor Heawood,” thinks Kempe. “He is now hooked on it, as I was fifteen years ago. He’d better drop it; this four colors thing is old hat.”

场景 3

Scene 3

2013 年 11 月 8 日下午 3 点，我坐在演讲厅后面。4^大约有三百名本科生参加这个星期五下午的“信息、计算和通信”课，这门课旨在向未来的土木工程师和机械工程师灌输（传达？）计算机科学的基础概念。我看到我的弟弟和他的朋友们——好学生——坐在第二排。他们刚刚开始他们的学术课程；我的课程快完成了。但我们却在同一个教室里，等待同样的信息（命令？）。教授调整了一下麦克风：“好的。大家好。上周我们讨论了奈奎斯特-香农采样定理。今天，我们将从克劳德·香农对数字信号的数学理解的另一个贡献开始，即香农-哈特利定理。这是一个非常强大的定理，可以用以下公式总结：

November 8, 2013, 3 p.m. I sit at the back of the lecture hall.⁴ Around three hundred undergraduate students are also attending this Friday afternoon “Information, Computing and Communication” class that aims to inculcate (communicate?) the foundational concepts of computer science to future civil and mechanical engineers. I see my younger brother and his friends—good students—in the second row. They’ve just started their academic curriculum; I’ve almost finished mine. But here we are in the same classroom, waiting for the same information (orders?). The professor adjusts his microphone: “All right. Hi, everyone. So, last week we talked about the Nyquist-Shannon sampling theorem. Today, we’ll start with another contribution of Claude Shannon to the mathematical understanding of digital signals, which is the Shannon-Hartley theorem. It is quite a powerful theorem that can be summarized with this formula here:

当然，我们会一起渡过难关。”

Of course, we’ll go through it together.”

此时，我们不需要对“定理”（场景 1 和 3）、“猜想”（场景 2）、“证明”（场景 1 和 2）和“公式”（场景 3）做出任何先验区分。我们只需要注意到，所有三个场景虽然可能与数学有关，但所涉及的主张或多或少都吸引了人们的支持。在场景 1 中，Elkan 关于模糊逻辑的主张首先引起了第十一届全国人工智能大会程序委员会的支持。但是，随后在 1994 年 1 月，他的主张遭到了许多逻辑学家同事的反驳，他们毫不犹豫地在网络论坛“comp.ai.fuzzy”上发表“反诉”。在场景 2 中，Kempe 关于 Francis Guthrie 的主张（“四色猜想”）的真实性的主张也首先引起了《美国数学杂志》编辑委员会的支持。但 1890 年夏天，肯普放弃了自己的主张并坚持希伍德的主张。然而，格思里 1852 年的“坦率”主张仍未失去其令人信服的力量，这让肯普对希伍德的命运感到疑惑。第三幕非常简单：香农和哈特利的主张——以及投射在演讲厅白板上的相关公式——即将向一群工程学本科生讲授。这里几乎没有疑问的余地：2013 年 11 月，香农和哈特利的主张吸引了相当多的人的支持。事实上，他们的主张如此强烈，以至于一种众所周知的教学手段——考试——很快就会验证所有学生是否正确地遵守了他们的主张。

At this point, we do not need to make any a priori distinction between “theorems” (scenes 1 and 3), “conjectures” (scene 2), “proofs” (scene 1 and 2), and “formulas” (scene 3). We just need to notice that all three scenes, while presumably concerning mathematics, deal with claims that attract more or less adherence. In scene 1, Elkan’s claim about fuzzy logic first attracts the adherence of the Eleventh National Conference on Artificial Intelligence’s program committee. But then, in January 1994, his claim repulses many logician colleagues who do not hesitate to publish “counterclaims” on the web forum “comp.ai.fuzzy.” In scene 2, Kempe’s claim about the veracity of Francis Guthrie’s claim (the “four colors conjecture”) also first attracts the adherence of the editorial board of the American Journal of Mathematics. But then, in the summer of 1890, Kempe dissociates himself from his own claim and adheres to that of Heawood. However, Guthrie’s 1852 “candid” claim has not lost all of its conviction strength yet, which makes Kempe puzzled about the fate of Heawood. Scene 3 is quite straightforward: Shannon and Hartley’s claim—and its correlated formula projected on the lecture hall’s whiteboard—is about to be taught to a crowd of undergraduate students in engineering. There is little room for doubt here: in November 2013, Shannon and Hartley’s claim attracts the adherence of quite a lot of people. In fact, their claim is so strong that a well-known pedagogical device—the exam—will soon verify that all students properly adhere to it.

这些基本但公正的观察是我们开始操作化练习所需要的全部。数学家当然做了很多事情，但在这些事情中，他们提出的主张吸引了更多或更少的人的追随者。那么让我们假设“定理”、“猜想”、“公式”或“证明”这些宏大的概念都可以以一种脚踏实地的方式掌握；让我们假设，在某种程度上，它们是说服更多或更少人的主张。

These basic but fair observations are all we need to start our operationalization exercise. Mathematicians certainly do a lot of things, but among these things, they make claims that attract the adherence of more or fewer individuals. Let us assume then that the grand notions of “theorems,” “conjectures,” “formulas,” or “proofs” can all be grasped in a down-to-earth manner; let us assume that, to a certain extent, they are claims that convince more or fewer individuals.

这种将数学知识（定理、猜想、证明、公式）视为某种修辞的产物的方式乍一听可能有些奇怪。许多宏大的叙事确实高呼数学真理的抽象力量，而这些真理本身据说描述了某种更高级的现实。5^但这恰恰是我们不想走的路，至少现在还不想。如果我们不想撞上数学认识论的尖锐岩石，我们就需要捂住耳朵，暂时忽略必然性的警报。幸运的是，我们的第一个操作假设——数学家提出的主张能说服更多或更少的人——很好地呼应了拉卡托斯 (1976) 的重要数学著作的核心论点。正如他所表明的那样，数学不应被视为不言而喻的发现的积累，而应被视为一个创造性的过程，在此过程中，并发的主张会受到批评和改进。但这些主张是如何受到批评或改进的呢？它们如何获得或失去相对的说服力？香农和哈特利在第 3 场中的主张似乎比埃尔坎在第 1 场中的主张更有力。同样，在 1890 年，肯普在 1879 年提出的主张在希伍德（第 2 场）的主张面前也显得无能为力。这种差异是如何产生的？

This way to consider mathematical knowledge—theorems, conjectures, proofs, formulas—as the product of some rhetoric may sound odd at first. Many grand narratives have indeed chanted the abstract power of mathematical truths that, by themselves, supposedly describe some superior reality.⁵ But this is precisely the road we do not want to take, at least not yet. If we do not want to crash on the sharp rocks of epistemological accounts of mathematics, we need to plug our ears and, for the moment, ignore the sirens of necessity. Fortunately for us, our first operational hypothesis—mathematicians make claims that convince more or fewer individuals—echoes well the central thesis of Lakatos’s (1976) important book on mathematics. As he showed, instead of an accumulation of self-evident discoveries, mathematics should be considered a creative process during which concurrent claims are subjected to criticism and improvement. But how are such claims criticized or improved? How do they gain or lose their relative conviction strength? Shannon and Hartley’s claim in scene 3 seems much stronger than Elkan’s claim in scene 1. Similarly, in 1890, the claim Kempe made in 1879 is now powerless in front of Heawood’s claim (scene 2). How do such differences come about?

为了更好地理解（数学）主张如何获得或失去说服力，我们需要对场景 1、2 和 3 进行另一个基本观察。如果更多或更少的人能够坚持场景中的主张，则意味着他们可以访问这些主张。什么媒介允许这种访问？有些主张是口头的，但我们显然不在这里处理它们。场景 1、2 和 3 中的主张都是书面的。这一重要特征使个人能够阅读它们并最终（极少数情况下）遵守它们。在场景 1 中，正是 Elkan 在会议记录中出现的书面主张使得程序委员会遵守它。但在 1994 年 1 月，正是网络论坛“comp.ai.fuzzy”上书面反诉的增加开始折磨 Elkan。在场景 2 中，Kempe 和 Heawood 都通过阅读数学期刊来访问他们各自的主张。最后，场景 3 中的工程专业学生被要求遵守投影在教室白板上的 Shannon 和 Hartley 的主张。当然，Shannon 和 Hartley 并没有在投影文件上写下他们的主张；许多人介入，将他们的主张穿越时空，直到到达这个特定的演讲厅。但这种翻译过程并没有改变主张的整体形状；它仍然是写在平面上的东西。因此，在这一点上，我们可以稍微更新一下我们的第一个假设：数学家确实做了很多事情，但在这些事情中，他们写的主张吸引了更多或更少的人的追随者。

To better understand how (mathematical) claims gain or lose conviction strength, we need to make another basic observation about scenes 1, 2, and 3. If more or fewer individuals could adhere to the scenes’ claims, it means that they could access these claims. What medium allowed such access? Some claims are oral, but we are obviously not dealing with them here. The claims in scenes 1, 2, and 3 are all written. This important characteristic allows individuals to read them and eventually—very rarely—adhere to them. In scene 1, it is Elkan’s written claim as it appears in the conference’s proceedings that makes the program committee adhere to it. But in January 1994, it is the multiplication of written counterclaims on the web forum “comp.ai.fuzzy” that begins tormenting Elkan. In scene 2, both Kempe and Heawood access their respective claims by reading mathematical journals. Finally, the engineering students in scene 3 are asked to adhere to Shannon and Hartley’s claim projected on the classroom’s whiteboard. Of course, Shannon and Hartley did not write their claim on the projected document; many individuals intervened to carry their claim further through time and space until reaching this specific lecture hall. But this translation process does not change the overall shape of the claim; it is still something that is written down on a flat surface. At this point, we can therefore slightly refresh our first hypothesis: mathematicians surely do a lot of things, but among these things, they write claims that attract the adherence of more or fewer individuals.

也可以公平地假设上述场景中的书面声明并非凭空出现。为了在会议论文集、专业网络论坛、数学期刊或计算机科学教授的幻灯片中发表，它们都必须通过一系列测试，这些测试决定了它们作为书面声明的存在。我同意这一假设与有影响力但有争议的思想家提出的生存形而上学有关——接近“过程思维”（参见引言）。那么让我们将其视为我们操作化练习所需的假设。“凡是经得起考验的都是真实的”（Latour 1993a）。上述（数学）书面声明是真实的；因此它们经受住了考验。但是什么考验？

It is also fair to assume that the written claims in the above scenes did not appear ex nihilo. In order to be published in proceedings, specialized web forums, mathematical journals, or the slides of a computer science professor, they all had to overcome a series of tests, trials upon which their existence as written claims depended. I agree that this hypothesis flirts with the metaphysics of subsistence—close to “process thought” (cf. introduction)—as proposed by influential, yet contested, thinkers. Let us then consider it an assumption we need for our operationalization exercise. “Whatever resists trials is real” (Latour 1993a). The above (mathematical) written claims are real; they thus resisted trials. But what trials?

经受考验，成为事实

Resisting Trials, Becoming Facts

关于（数学）书面主张（如场景 1、2 和 3 中的主张）的说服力，我们可以考虑的第一种考验是它们在实际发表之前必须经历的考验。检查我们通常所说的主张的“来源”确实是评估其严肃性的常用方法。

The first kind of trial we can consider regarding the conviction strengths of (mathematical) written claims such as those in scenes 1, 2, and 3 are the trials they must endure before their actual publication. Examining what we often call the “sources” of claims is indeed a common way to evaluate their seriousness.

例如，我们可以合理地假设，在其他条件相同的情况下，发表在《自然》杂志上的主张通常比发布在几乎没有监控的社交媒体平台上的主张更有说服力。即使不考虑它们各自的内容，这两个主张的能力也是不同的。为什么会这样？我们必须立即把声望或象征权力的问题放在一边；这些是我们社会学研究方法禁止操纵的捷径。对这个话题进行更实证的把握很快就会指出有多少人可以阻止发表一项主张。很少有人——或者机器人——能阻止我在 Facebook 等平台上发表主张。相反，许多人可以阻止我在《自然》杂志上发表主张。考虑到那些必须被主张说服的人，这些主张才能传播并接触到更广泛的受众，这一点至关重要，因为它在某种程度上衡量了分歧的成本。如果有人不同意我在 Facebook 上发表的主张，他们可以耸耸肩，转而去做别的事情。⁶但如果同一个人不同意我在《自然》上发表的某个观点，他们就必须不同意我、我的机构、支持我研究的资助机构、《​​自然》的编辑委员会、负责提名该委员会的人等等的观点。与我在 Facebook 上发表的观点相比，我在《自然》上发表的观点最初得到了更大范围的外部盟友团队的支持（Latour 1987，31–33）。

For example, we can make the fair assumption that, all things being equal, a claim published in the journal Nature will generally have more conviction strength than a claim posted on some social media platform with very little monitoring. Without even considering their respective content, both claims will have different capabilities. Why is that? We must immediately put aside the question of prestige or symbolic power; these are shortcuts our sociological method of inquiry forbids us to manipulate. A more empirical grip on this topic would quickly point to the number of individuals who could prevent the publication of a claim. Very few people—or bots—can prevent me from publishing a claim on, say, Facebook. Conversely, many individuals can prevent me from publishing a claim in the journal Nature. Taking into account those who have to be convinced by claims in order for them to circulate and reach a broader audience is crucial as it somewhat calibrates the cost of disagreement. If someone disagrees with a claim I publish on Facebook, they can just shrug their shoulders and move on to something else.⁶ But if the same person disagrees with a claim I publish in Nature, they will have to disagree with me, my institution, the funding agencies that supported my research, Nature’s editorial board, those responsible for the nomination of this board, and so on. Compared with a claim I publish on Facebook, a claim I publish in Nature is initially supported by a far bigger team of external allies (Latour 1987, 31–33).

但是，如果我们考虑一下这三个场景，我们很快就会意识到，通过出版审判——从而获得外部盟友——并不足以保证（数学）主张具有持久的说服力。虽然从坚定的守门人的角度来看，这次讲座可能足以迅速解释香农和哈特利的主张在演讲厅内的说服力——学生们被所有外部盟友（他们的教授、他们的手册、所有负责他们大学工程课程的人、他们很快就要通过的考试）彻底击败——但这并不能帮助我们理解肯普、希伍德和埃尔坎的主张的相对强度（场景 1 和场景 2）。在场景 2 中，肯普和希伍德的主张都通过了类似的出版审判；这两个主张最初都得到了大致相同数量的人的支持。7^然而，随着希伍德的主张似乎得到证实，肯普的主张开始变得不可信。场景 1 的情况更加令人困惑：尽管埃尔坎的主张接连经受住了负责出版诉讼程序和评选“最佳写作奖”的 68 名个人的审查，但论文”，⁸他的主张被几乎没有监控的网络论坛上的帖子严重动摇（Rosental 2003，81-86）。同样，这些反诉必须经受住其他类型的考验才能获得这样的力量。

But if we consider our three scenes, we quickly realize that surviving publication trials—and thus enrolling external allies—is not enough to assure any durable conviction strength of (mathematical) claims. Although this lecture, in terms of convinced gatekeepers, may be enough to quickly account for the conviction strength of Shannon and Hartley’s claim within the lecture hall—the students being literally crushed by all its external allies (their professor, their manuals, all those responsible for the engineering curriculum of their university, the exam they will soon have to pass)—it does not help us understand the relative strengths of Kempe’s, Heawood’s, and Elkan’s claims (scene 1 and scene 2). In scene 2, both Kempe’s and Heawood’s claims survived similar publication trials; both propositions were initially supported by roughly the same number of individuals.⁷ Yet Kempe’s claim became distrusted as Heawood’s appeared certified. The situation is even more confusing in scene 1: even though Elkan’s claim successively resisted the scrutiny of the sixty-eight individuals responsible for the publication of the proceedings and the selection of the “Best Written Paper,”⁸ his claim is seriously shaken up by posts on a web forum with almost no monitoring (Rosental 2003, 81–86). Again, these counterclaims must have survived other kinds of trials in order to gain such strength.

另一种可能为书面主张提供力量的审判是通过引用和参考文献的方式连续招募内部盟友（Latour 1987，33-45）。用先前发表的主张来充实自己的主张确实是一种重要的说服策略，甚至已经成为一个完整的研究领域。9^除了书面文件之外的盟友之外，现在有参考文献和引文的主张也得到了其内部盟友的支持。真的是这样吗？虽然经常需要，但通过参考文献和引文来增强主张的说服力可能是一项冒险的尝试。如果参考文献与主张不符，或者更糟的是，如果一些未提及的参考文献与提出的主张相矛盾，该怎么办？在某些情况下，这种引用审判被克服了。一个例子是香农的初始论文，该论文提出了后来被称为“香农-哈特利定理”的基本要素（香农 1948）。在这篇论文中，香农引用了拉尔夫·哈特利（因此后来被纳入该定理的名称）和其他 13 位重要数学家先前提出的“已固化”的论断。据我所知，在香农首次发表论文后，并没有出现关于这些参考文献使用的严重分歧。但埃尔坎的论文却并非如此。尽管他动员了 39 位内部盟友来支持他关于模糊逻辑局限性的论断，但他的反对者设法在专门的网络论坛上找到并发表了许多强有力的“反驳”。埃尔坎很快就被发现不知道模糊逻辑在高级专家系统中的许多最新用途（Rosental 2003，157-168）。尽管他们最初确实有助于说服第十一届全国人工智能大会的程序委员会，但埃尔坎论文的内部盟友最终却成为反对者的垫脚石。

Another kind of trial that may provide strength to written claims is one that consists in successively enrolling internal allies by means of citations and references (Latour 1987, 33–45). Equipping one’s claim with previously published claims is indeed an important conviction strategy that has even become a whole field of study.⁹ In addition to allies outside of the written document, a claim with references and citations is now supported by allies inside of it. Or is it? While often necessary, augmenting the conviction strength of a claim by means of references and citations can be a risky endeavor. What if the references do not match the claim, or worse, what if some unmentioned references contradict the presented claim? In some cases, this citation trial is overcome. One example is Shannon’s initial paper that presented the basic elements of what would later be called the “Shannon-Hartley theorem” (Shannon 1948). In this paper, Shannon enrolls previously “solidified” claims made by Ralph Hartley (hence his later inclusion in the theorem’s name) and thirteen other important mathematicians. As far as I know, no serious disagreements about the use of these references emerged after Shannon’s initial publication. But the same was not true of Elkan’s publication. Although he mobilized thirty-nine internal allies to strengthen his claim about the limitations of fuzzy logic, his contradictors managed to find and publish many strong “counter references” on the specialized web forum. Elkan soon appeared as someone unaware of many recent uses of fuzzy logic in advanced expert systems (Rosental 2003, 157–168). Although they were at first certainly useful to convince the program committee of the Eleventh National Conference on Artificial Intelligence, the internal allies of Elkan’s paper ended up working as stepping stones for his contradictors.

然而，是否能经受住引证审判，同样不足以解释我们所有场景中主张的相对说服力。事实上，在场景 2 中，肯普 1879 年的论文只三次提到了以前的数学命题，前两次是奥古斯都·德·摩根和亚瑟·凯莱对伦敦数学学会的松散陈述（肯普 1879，193-194），第三次是奥古斯丁-路易·柯西关于多面体的更重要的主张（肯普 1879，198）。然而，这种引用的稀缺并没有阻止他的主张——证明格思里 1852 年的命题是正确的——令人信服。数学家同事们已经争论了十一年。希伍德的主张更是如此，因为他 1890 年的论文除了引用肯普 1879 年的论文外没有其他参考文献。同样，这种稀缺性并没有阻止他的主张吸引主要相关人士：肯普本人的拥护（MacKenzie 1999, 22）。发表的（数学）主张中一定有其他东西，有时可以使它们具有说服力。

However, surviving or not surviving citation trials is, again, not enough to account for the relative conviction strengths of the claims in all of our scenes. Indeed, in scene 2, Kempe’s 1879 paper makes only three references to former mathematical propositions, the first two being loose statements made by Augustus De Morgan and Arthur Cayley to the London Mathematical Society (Kempe 1879, 193–194) and the third one being a more important claim made by Augustin-Louis Cauchy about polyhedrons (Kempe 1879, 198). Yet this scarcity of references did not prevent his claim—the proof that Guthrie’s 1852 proposition was correct—from convincing his mathematician colleagues for eleven years. The same is even truer of Heawood’s claim, for his 1890 paper includes no references other than Kempe’s 1879 paper. Again, this scarcity did not prevent his claim from attracting the adherence of the chief person concerned: Kempe himself (MacKenzie 1999, 22). There must be something else in published (mathematical) claims that makes them gain, sometimes, in persuasion strength.

一些潜在的反对已发表的（数学）主张的人不会对作者列举的坚定看门人或引用的参考文献留下深刻印象。为了被一个主张说服，这些持怀疑态度的读者希望看到作者要求他们相信的东西。这种向读者展示所讨论事物的策略正是希伍德在反对肯普的论文中使用的策略。他不仅依靠外部盟友；他还展示了一个图形（见图5.1），根据肯普 1871 年的主张，这个图形是不可能画出来的：

Some potential objectors of published (mathematical) claims will not be impressed by lists of convinced gatekeepers nor by the references invoked by the author. To be convinced by a claim, these skeptical readers want to see the thing the author asks them to believe in. This strategy that consists of presenting the thing in question to the reader was precisely the one used by Heawood in his paper against Kempe. He did not only rely on external allies; he also showed a figure (see figure 5.1) that, according to Kempe’s 1871 claim, was impossible to draw:

肯普先生说——在 E 的红绿区域和 B 的红黄区域中传输颜色将各自去除一种红色，所需的工作就完成了。如果是这样，这将立即证明所讨论的命题 [四色猜想]。……但不幸的是，可以想象，虽然任何一种转置都会去除一种红色，但两者可能无法去除两种红色。图 [下面] 是这种可能性的实际例证。（Heawood 1890，337–338）

Mr. Kempe says—the transmission of colours throughout E’s red-green and B’s red-yellow regions will each remove a red, and what is required is done. If this were so, it would at once lead to a proof of the proposition in question [the four-colours conjecture]. … But, unfortunately, it is conceivable that though either transposition would remove a red, both may not remove both reds. Fig [below] is an actual exemplification of this possibility. (Heawood 1890, 337–338)

我们不需要花太多时间在希伍德图形¹⁰的特殊性上，也不需要花太多时间在已发表的数学主张中附图的作用上。11这里，重要的是要注意信念策略；就像从事许多其他领域（生物学（Rheinberger ¹⁹⁹⁷）、化学（Bensaude-Vincent 1995）、气候学（Edwards 2013））的科学家一样，数学家试图通过添加他们所写内容的指称来增强说服力。那么，在这一点上，“这不再是一个信仰问题：这是一个观察问题”（Latour 1987, 48）。如果到目前为止，我将形容词“数学”放在括号中，那并不是要赋予数学主张过多的特殊性；它们也是试图通过聚集越来越多的支持者来压制潜在反对者的科学流派的一部分。科学和数学文本确实可以比作雪橇轨道，几乎没有回旋余地，但需要很高的技巧。在这两种情况下，读者都必须从 A 点开始，经过检查点 B _1,2, … ,n，最后到达 C 点，即试图确立为事实的主张。

We do not need to spend too much time on the specificities of Heawood’s figure¹⁰ nor on the role of drawings in published mathematical claims.¹¹ Here, the important thing to notice is the conviction strategy; just as scientists engaged in many other fields—biology (Rheinberger 1997), chemistry (Bensaude-Vincent 1995), climatology (Edwards 2013)—mathematicians try to gain in persuasion strength by adding the referent of what they write about. At this point, then, “this is not a question any more of belief: this is seeing” (Latour 1987, 48). If, until now, I put the adjective “mathematical” in parenthesis, it was not to grant too much specificity to mathematical claims; they too are part of the scientific genre that tries to silence potential objectors by gathering more and more supporters. Scientific as well as mathematical texts can indeed be compared with bobsled tracks allowing very little room for maneuver while implying high level of skills. In both cases, readers must start at point A, pass through checkpoints B_1,2,…,n, and finally finish at point C, the claim that tries to be established as a fact.

如果科学文献可以被描述为聚集许多外部和内部盟友的文本，以孤立其读者并迫使他们接受不同的科学领域逐渐形成了自己特定的修辞习惯。12^在数学中，整个captation 试验（Latour 1987, 56–61）在于巧妙地控制潜在反对者的动作，这一点已由 Rotman (1995, 2006) 进行了细致的分析。如他所示，数学出版物中充满了祈使句形式的动词，例如“构造”、“定义”、“连接”或“计算”。但仔细分析这些祈使句形式就会发现，它们实际上分为两种不同的类型：包容性祈使句，用于建立前提——通常配有参考文献；排他性祈使句，用于列出想象中的读者为达到所声称的结果应该执行的动作列表：

If scientific literature can be described as texts gathering many external and internal allies in order to isolate their readers and force them to take only one path, different scientific domains progressively shaped their own specific rhetorical habits.¹² In the case of mathematics, this whole captation trial (Latour 1987, 56–61) that consists in subtly controlling the movements of potential objectors has been finely analyzed by Rotman (1995, 2006). As he showed, mathematical publications are full of verbs in the imperative form, such as “construct,” “define,” “connect,” or “compute.” But a close analysis of these imperative forms reveals that they are in fact split into two distinctive types: inclusive imperative to establish premises—often equipped with references—and exclusive imperative to present lists of actions an imaginary reader should perform to reach the claimed result:

包容性命令以动词“考虑”、“定义”、“证明”及其同义词为标志，要求说话者和听话者建立并生活在一个共同的世界，或者他们对这个世界中的某件事有共同的特定争论信念；而排他性命令——本质上是所有其他动词表示的数学动作——要求执行在已经共享的世界中有意义的某些操作。（Rotman 2006，104）

Inclusive command marked by the verbs “consider,” “define,” “prove” and their synonyms—demand that speaker and hearer institute and inhabit a common world or that they share some specific argued conviction about an item in such a world; and exclusive commands—essentially the mathematical actions denoted by all other verbs—dictate that certain operations meaningful in an already shared world be executed. (Rotman 2006, 104)

这些元素对于我们的操作化练习至关重要，因为它们表明了数学文本中标题试验的恰当条件。如果持怀疑态度的读者在作者动员的所有盟友的帮助下别无选择，只能接受前提并遵循一条特定的路径以得出一个必要的结论，那么数学文本及其伴随的主张至少在暂时上克服了他们的标题试验。在这方面，肯普 1879 年关于四色猜想的论文非常具有说明性。请记住，肯普想要证明四种颜色足以为平面上绘制的任何地图着色，使得没有邻国具有相同的颜色。他是如何要求读者得出这个结论的？通过一系列包容性命令，肯普和他想象中的怀疑论读者都首先定义一个完美的四色“单连通表面”，并将其分为许多“区域”（肯普 1879，193）。一旦建立了这个基本的共同世界，他们就会考虑两组“分离区域”，要么用红色和绿色，要么用黄色和蓝色 (Kempe 1879, 194)。这些前提允许 Kempe 和他的读者进一步定义“汇合点”（边界和区域相交的点）的属性，这些属性本身允许定义具有不同特征的六类区域：“岛屿区域”、“岛屿区域”、“半岛区域”、“半岛区域”、“复杂区域”和“简单区域”（Kempe 1879, 195–196）。一旦建立了这个相当复杂的共同世界，已经制定，Kempe 随后切换到独占命令并要求他的读者执行一系列操作：

These elements are crucial for our operationalization exercise as they indicate the felicity conditions of captation trials within mathematical texts. If skeptical readers, thanks to all the allies mobilized by the writer, have no other choice than to accept the premises and follow one specific path in order to reach one necessary conclusion, a mathematical text and its concomitant claim have, at least temporally, overcome their captation trial. In this respect, Kempe’s 1879 paper on the four colors conjecture is quite illustrative. Remember that Kempe wanted to prove that four colors suffice to color any map drawn on a plane in such a way that no neighboring countries have the same color. How did he enjoin his readers to reach this conclusion? With a succession of inclusive commands, both Kempe and his imaginary skeptical reader start by defining a perfectly four-colored “singly connected surface” divided into many “districts” (Kempe 1879, 193). Once this basic common world has been instituted, they then consider two sets of “detached regions” either colored in red and green or in yellow and blue (Kempe 1879, 194). These premises allow Kempe and his reader to further define the properties of “points of concourse” (points where boundaries and districts meet) that themselves permit the definition of six classes of districts with different characteristics: “island districts,” “island regions,” “peninsula districts,” “peninsula regions,” “complex districts,” and “simple districts” (Kempe 1879, 195–196). Once this quite complex common world has been instituted, Kempe then switches to exclusive commands and asks his reader to execute a series of operations:

现在，拿出一张纸，将其剪成与任何简单岛屿或半岛区相同的形状，但要大一些，这样在放在区上时，边界刚好重叠。将这块补丁（我将这样称呼它）固定在表面上，并画出与补丁相交的所有边界……在补丁内的一点（交汇点）相交。如果只有两条边界与补丁相交，如果区是半岛，就会发生这种情况，则将它们穿过补丁连接起来，不需要交汇点。这样地图上就会少一个区，边界的数量也会减少。（Kempe 1879，196-197；斜体添加）

Now, take a piece of paper and cut it out to the same shape as any simple-island or peninsula-district, but larger, so as just to overlap the boundaries when laid on the district. Fasten this patch (as I shall term it) to the surface and produce all the boundaries which meet the patch … to meet at a point, (a point of concourse) within the patch. If only two boundaries meet the patch, which will happen if the district be a peninsula, join them across the patch, no point of concourse being necessary. The map will then have one district less, and the number of boundaries will also be reduced. (Kempe 1879, 196–197; italics added)

通过要求读者重复这一修补过程，整个想象中的地图逐渐缩小为一个没有边界或汇合点的单一区域。然后，肯普要求读者逆转这一过程；也就是说，“以相反的顺序剥去补丁，先剥掉最后贴上的补丁。随着每个补丁的剥落，它露出了一个新的区域，地图逐渐发展起来”（肯普 1879，197）。就在这时，肯普再次切换到包容性命令，从而在刚刚修改的第一个世界的基础上建立了第二个共同世界。作者和读者再次一起定义了所有区域、边界和汇合点的逐步重建。渐渐地，他们很快意识到，他们对区域、边界和会聚点的重新组合分别相当于奥古斯丁-路易·柯西在 1813 年定义的多面体的面、边和点（Kempe 1879, 198）。一旦建立了这个多面体世界，Kempe 最后一次切换到独占命令，让读者达到声称的结果：显然——看，我们刚刚一起做到了！——四种颜色足以给平面上绘制的任何地图着色，使得没有邻国具有相同的颜色。

By asking the reader to reiterate this patching process, the whole imagined map is progressively reduced to one single district with no boundaries or points of concourse. Kempe then asks the reader to reverse the process; that is, to “strip off the patches in reverse order, taking off first that which was put on last. As each patch is stripped off it discloses a new district and the map is developed by degrees” (Kempe 1879, 197). At this precise point, Kempe switches to inclusive command again, thus instituting a second common world based on the first one that has just been modified. The author and the reader, together again, define the progressive reconstitution of all districts, boundaries, and points of concourse. Little by little, they soon realize that their recombination of districts, boundaries, and points of concourse is equivalent to, respectively, faces, edges, and points of polyhedrons as already defined by Augustin-Louis Cauchy in 1813 (Kempe 1879, 198). Once this polyhedron world has been instituted, Kempe switches one last time to exclusive command and makes the reader reach the claimed result: obviously—look, we have just done it together!—four colors suffice to color any map drawn on a plane in such a way that no neighboring countries have the same color.

我们不需要理解肯普论文的每一个小步骤。我们只需要欣赏肯普如何设法控制读者的动作；从最初的前提到结论，读者实际上被肯普的论证线索所引导。他的盟友非常多——“单连通表面”、“区域”、“分离区域”、柯西的“多面体”——他的过渡足够流畅，可以引导读者穿越必然性的流动。但正如我们所见，肯普的俘虏只是暂时的，因为十一年后，希伍德设法摆脱了肯普的论证线索，并提出了一个摧毁整个修辞大厦的形象（见图5.1）。

We do not need to understand every little step of Kempe’s paper. We just need to appreciate how Kempe manages to control the movements of his reader; from the initial premises to the conclusion, the reader is literally carried through Kempe’s line of argument. His allies are quite numerous—“single connected surface,” “districts,” “detached regions,” Cauchy’s “polyhedrons”—and his transitions are smooth enough to transport the reader through the flow of necessity. But as we saw, Kempe’s captatio was only temporary, for as eleven years later, Heawood managed to escape from Kempe’s line of argument and propose a figure that dismantled the whole rhetorical edifice (see figure 5.1).

出版、引用和标题审判——就像任何其他试图获得说服力并成为事实的主张一样，数学主张必须经受住许多危险。但这还不够。一个发表在重要期刊上的主张，有详尽的参考文献和流畅的论证思路，如果后续主张没有进一步发展，它仍可能消失。这是一个必要条件，因为没有单独的科学事实：“事实构建是一个集体过程，孤立的人只会构建梦想、主张和感觉，而不是事实”（Latour 1987, 41）。主张的命运，即它逐渐转变为凝固的事实，最终取决于后续主张如何使用它。我们看到，尽管肯普的主张具有标题强度，但最终还是被希伍德驳斥了。从数学事实的地位来看，它变成了纯粹的虚构。希伍德的主张呢？很难称其为事实，因为它只涉及肯普的虚构；它接连驳斥了肯普的主张，但没有提供任何可证实或可反驳的命题。那么埃尔坎的主张又如何呢？尽管埃尔坎努力使其更有力——尤其是通过纳入许多合著者、更好地排列参考文献和更流畅的过渡（Elkan et al. 1994; Rosental 2003, 282–331）——但它最终因引起的质疑而出名；也就是说，恰恰因为它不是事实。在我们任意的数学例子中，只有香农的主张经受住了这一重要的后人考验，正如场景 3 已经暗示的那样。事实上，香农的主张在后人考验中表现得如此出色，以至于它逐渐成为极少数事实的一部分，这些事实不断被用作后来主张的资源。随着它越来越多地被纳入而没有任何怀疑方式，它变成了一个黑匣子，其中的认证内容以明确的形式呈现。这种程式化过程（Latour 1987, 42）是科学事实在后来的主张中被广泛采用的典型过程。尽管香农在他的第一篇论文中进行了多次论证，但只有这些论证的结果被逐渐保留下来。这些结果后来被串联、完善，并与哈特利建立的先前结果相联系，直到达到场景 3 中提出的公式所表达的程式化形式。也许很快，这个强有力的数学事实甚至可能成为一个“单句陈述”（Latour 1987, 43）：一个被广泛接受以至于不再需要任何引用的科学事实。如果发生这种情况，香农和哈特利的定理将成为心照不宣、无可争辩和必要的知识的一部分。

Publication, citation, and captation trials—just as any other claim trying to gain conviction strength and become a fact, mathematical claims must survive many jeopardies. Yet this is still not enough. A claim published in an important journal, with well-arrayed references and a smooth line of argument, may still vanish if it is not carried further by later claims. This is a sine qua non condition as there is no such thing as a solitary scientific fact: “Fact construction is so much a collective process that an isolated person builds only dreams, claims and feelings, not facts” (Latour 1987, 41). The fate of a claim, its progressive transformation into a solidified fact, depends ultimately on how it is used by later claims. We saw that Kempe’s claim, despite its captation strength, ended up being refuted by Heawood. From the status of mathematical fact, it turned into mere fiction. What about Heawood’s claim? It is difficult to call it a fact as it only concerned Kempe’s fiction; it successively refuted Kempe’s claim but did not provide any confirmable, or refutable, proposition. What about Elkan’s claim, then? Despite Elkan’s efforts to make it stronger—especially via the inclusion of many coauthors, better arrayed references, and smoother transitions (Elkan et al. 1994; Rosental 2003, 282–331)—it ended up being known for the doubtful reactions it gave rise to; that is, precisely, for not being a fact. Among our arbitrary mathematical examples, only Shannon’s claim survived this important posterity trial, as scene 3 already suggested it. In fact, Shannon’s claim survived the posterity trial so well that it progressively became part of a very small number of facts that are constantly used as resources in later claims. As it became more and more enrolled without any skeptical modalities, it became a black box with certified content presented in a clear-cut form. This stylization process (Latour 1987, 42) is typical of scientific facts that are much enrolled in later claims. Although Shannon went through several demonstrations in his initial paper, only the results of these demonstrations were progressively retained. These results were later concatenated, polished, and linked with former results established by Hartley until reaching a stylized form expressed by the formula presented in scene 3. Soon, perhaps, this strong mathematical fact may even become a “single sentence statement” (Latour 1987, 43): a scientific fact that is so accepted that it no longer needs any reference. If this happens, Shannon and Hartley’s theorem will be part of tacit, undisputable, and necessary knowledge.

关于黑盒抛光事实可能成为隐性知识的一部分的最后这些要素使我们能够回应一个重要的反对意见：

These last elements about blackboxed polished facts that may become part of tacit knowledge allow us to respond to an important objection:

持怀疑态度的读者的反对意见

Objection of a skeptical reader

但是数学与其他科学学科的不同之处难道不在于它处理的是基本事实吗？当你介绍肯普的论文时，我们就能感受到这一点：为了克服标题审判，他遵循了永恒的演绎法则，不是吗？

But is not mathematics different from all the other scientific disciplines in that it deals with fundamental truths? We could feel it when you presented Kempe’s paper: in order to overcome the captation trial, he followed the timeless laws of deduction, did he not?

不久前，回应这一经典的反对意见还会非常困难。13^但由于 Reviel Netz (2003, 2004) 在语言学方面的努力，我们现在知道，我们所说的“演绎”和“逻辑关系”本身就是黑匣子里经过打磨的事实，最初于公元前5 世纪中叶在希腊和意​​大利南部发表。14^当时，几位自学成才的业余爱好者（大概是想与古希腊极具争论性的文化保持距离）¹⁵惊讶地发现，当他们只写关于绘制在蜡板上的字母图表的属性时，他们可以逐步表达无可争辩的命题。更准确地说，通过从图表的某些字母部分（比如两个线段）开始，他们可以依次将它们与同一图表的另一个字母部分进行比较。这种非常基本的操作是通过在平面上结合图画和字母实现的，可以重构为：“这里的线段 A 等于那里的线段 B。而那边的线段 B 等于那边的线段 C。”反过来，由于字母图，希腊几何学家可以偷偷地以必要的方式使用连接副词：“所以，这里的线段 A 等于那边的线段 C。”这种转变看似微不足道，但实际上至关重要。事实上，这个第一个必要结果可用于比较图的其他部分：“那边的线段 D 是线段 C 的两倍。因此，线段 A 是线段 D 的一半。”逐渐地，通过比较图的更多部分，使用越来越多的连接副词并累积越来越多的中间结果（例如“A 是线段 D 的一半”），希腊几何学家最终可以得到一个复杂但必然的真命题——从他的第一个基本断言到最后一个复杂断言的书面指示步骤列表就是他主张真实性的证明。

Not so long ago, it would have been very difficult to respond to this classical objection.¹³ But thanks to the philological efforts made by Reviel Netz (2003, 2004), we now know that what we call “deduction” and “logical relations” are themselves blackboxed polished facts that were initially published around the middle of the fifth century BCE in Greece and southern Italy.¹⁴ At that time, several self-educated amateurs who, presumably, tried to distance themselves from ancient Greece’s highly polemical culture,¹⁵ were surprised to discover that when they wrote only about the properties of lettered diagrams drawn on wax tablets, they could, step by step, express indisputable propositions. More precisely, by starting with some lettered parts of a diagram—say, two segments—they could, in turn, compare them with another lettered part of the same diagram. This very basic operation, made possible by the combination of drawings and letters on a flat surface, can be reconstituted as such: “This segment A here is equal to that segment B there. And that segment B there is equal to that segment C over there.” In turn, thanks to the lettered diagram, Greek geometers could surreptitiously use conjunctive adverbs in a necessary way: “Therefore this segment A here is equal to that segment C over there.” The shift seems trivial but is in fact crucial. Indeed, this first necessary result could be used to compare other parts of the diagram: “And that segment D over there is two times segment C. Therefore, segment A is half segment D.” Progressively, by comparing more and more parts of the diagram, using more and more conjunctive adverbs and cumulating more and more intermediary results such as “A is half segment D,” the Greek geometer could end up with a complicated yet necessary true proposition—the written list of indexical steps going from his first basic assertion to his last complicated one being the proof of the veracity of his claim.

本节仅尝试将数学主张作为更广泛的科学主张的一部分，因此我们无需进一步深入研究 Netz 所做的令人着迷的工作。这里只需说，由于他的努力，我们现在可以有信心地断言，即使是演绎是过去被接受的主张的固化产物。这些构造出来但完全合乎逻辑的必然性定律在古希腊一定令人惊讶。16^但经过几个世纪的进一步主张，这种推理方式——显然经受住了后人的考验——逐渐被黑箱化、打磨和风格化，直至获得无可争辩的知识地位。17现在谁会在使用肯定前件推理规则时引用亚里士多德？然而，即使是这些逻辑原理——形式主义数学流派所珍视的^{18 ——}也经历了与香农和哈特利定理类似的过程，现在很少有信号处理数学家会试图反驳这个定理。正如它们帮助塑造的定理一样，演绎定律本身很久以前也是由配备特定工具的人塑造的（在这种情况下，是画在蜡板上并索引到小希腊句子的字母图表）。

For the sake of this section that only tries to present mathematical claims as part of the broader family of scientific claims, we do not need to dig further into the fascinating work done by Netz. Suffice it here to say that thanks to his efforts, we can now assert with some confidence that even deduction is the solidified product of past accepted claims. These constructed-yet-fully-logical laws of necessity must certainly have been surprising in ancient Greece.¹⁶ But after centuries of enrollments in further claims, this style of reasoning—that obviously overcame its posterity trial—was progressively blackboxed, polished, and stylized until acquiring the status of indisputable knowledge.¹⁷ Who would now quote Aristotle when using the inference rule of modus ponens? Yet even these principles of logic—dear to the formalist school of mathematics¹⁸—went through a process similar to that of Shannon and Hartley’s theorem that very few mathematicians in signal processing would now try to contest. Just as the theorem they helped to shape, deductive laws were themselves shaped a long time ago by people equipped with specific instruments (in this case, lettered diagrams drawn on wax tablets and indexed to small Greek sentences).

平板实验室

Flat Laboratories

在前面的部分中，我们花了一些时间来强调数学和科学主张之间的相似之处。看来两者都需要经历类似的考验才能最终成为无可争辩的事实。没有任何更高的必要性可以帮助数学主张成为经过认证的事实；它们也需要说服读者，以便加入到以后的主张中，并成为（极少数情况下）精致的黑匣子。

In the previous sections, we spent some time trying to stress the similarities between mathematical and scientific claims. It appeared that both need to survive similar trials to become, eventually, indisputable facts. No superior necessity helps mathematical claims to become certified facts; they too need to convince their readers in order to be enrolled in later claims and become, very rarely, polished black boxes.

然而，到目前为止，我们只考虑了硬币的一面。虽然研究数学出版的主张有助于我们认识到成功的数学命题可以被视为真正的认证知识，但我们可以合理地假设数学家不会把所有工作时间都花在准备、撰写和阅读论文上。他们还必须花时间和精力在他们所写的东西上。我们在上一节中考虑的所有主张确实都是关于事物的：Elkan 的模糊逻辑系统的局限性、Kempe 的四色猜想、Heawood 的 Kempe 关于四色猜想的主张以及 Shannon（以及后来的 Hartley）的噪声信道上信息传输的最大速率。但这些东西是如何组装起来的？哪些实践导致这些数学事物（或对象）在出版的材料中呈现？这些实践与其他科学界的实验室实践有何不同？

However, so far, we have only considered one side of the coin. Although looking at mathematical published claims helps us realize that successful mathematical propositions could be considered genuine certified knowledge, we can legitimately assume that mathematicians do not prepare, write, and read papers all their working time. They must also spend time and energy on the things they write about. All the claims we considered in the last sections were indeed about things: limitations of fuzzy logic systems for Elkan, the four colors conjecture for Kempe, Kempe’s claim about the four colors conjecture for Heawood, and maximum rate of information transmission over noisy channels for Shannon (and later, Hartley). But how are these things assembled? What practices lead to the presentation of these mathematical things—or objects—in published materials? Are these practices different from laboratory practices in other scientific communities?

当我们准备探究数学对象形成的位置时，我们立即面临一个困难：很少有经验的对这些位置的研究。尽管对数学领域内的争议进行了大量的研究（Warwick 1992、1993；MacKenzie 1999、2000、2004、2006；Rosental 2003、2004），并从著名数学家的日志中重建了数学对象形成的历史（Lakatos 1976；Pickering and Stephanides 1992），但对数学的实验室研究却很少。19^因此，在有限的资源下，我现在将尝试更多地强调数学的科学性：

As we prepare to look inside the locations in which mathematical objects are shaped, we immediately face a difficulty: there are very few empirical studies of such locations. Although there are robust studies about controversies within mathematical domains (Warwick 1992, 1993; MacKenzie 1999, 2000, 2004, 2006; Rosental 2003, 2004) and historical reconstructions of the shaping of mathematical objects from famous mathematicians’ logbooks (Lakatos 1976; Pickering and Stephanides 1992), there are very few laboratory studies of mathematics.¹⁹ It is thus with limited means that I will now try to stress the scientific aspect of mathematics a little bit more:

场景 4

Scene 4

1972 年冬天，加利福尼亚州拉霍亚市索尔克生物研究所。20 Paul Brazeau 很紧张。他的老板 Roger Guillemin 教授正在追捕他，怀疑他能否处理实验室全新的、非常昂贵的放射免疫^测定法。确实，最近从大型生物电子仪器打印出来的图表令人惊讶；它没有显示 Guillemin 新纯化的肽会触发生长激素，而是显示它会降低生长激素。这让 Guillemin 抓狂不已。但 Brazeau 和他的技术人员对整个实验过程进行了十几次回溯检查：没有出现任何错误。在精心组装的大鼠垂体细胞培养物中注射了适量的纯化肽，并且在放射免疫测定操作过程中没有发生任何操作失误。“这非常简单，”Brazeau 想。“要么我不是尽职尽责的专业人士，要么在过去三年里，我们对这种肽的判断都是错误的。”

Salk Institute for Biological Studies at La Jolla (California), winter of 1972.²⁰ Paul Brazeau is on edge. His boss, Professor Roger Guillemin, is after him, casting doubts on his ability to handle the lab’s brand new—and very expansive—radioimmunoassay. It is true that the graphs recently printed by the massive bioelectronic instrument are surprising; instead of showing that Guillemin’s newly purified peptide triggers the growth hormone, it shows that it decreases it. This drives Guillemin crazy. But Brazeau and his technicians retro-inspected the whole experimental procedure a dozen times: there were no mistakes. The right amount of purified peptide was injected in the carefully assembled rat pituitary cell culture, and no mishandling occurred during the operationalization of the radioimmunoassay. “It’s terribly simple,” thinks Brazeau. “Either I am no conscientious professional or, for the last three years, we were all wrong about this peptide.”

场景 5

Scene 5

都柏林，1843 年秋。威廉·罗文·汉密尔顿心情激昂：尽管他在将复数理论扩展到三维空间的尝试中再次陷入僵局，但他显然取得了重大进展。21^他对自己的新起点特别自豪；从令人厌倦的代数模型开始他之前的实验真是一个错误！现在他从几何上开始，从x + iy移动到x + iy + jz，他拥有一个更容易测试的三维线段（即使它从一开始就增加了第二个虚数j）。从这个意义上说，他的第一次实验非常具有决定性。多亏了他的德国同事戈特霍尔德·艾森斯坦的建议，他可以通过放弃假设i和j之间可交换。然后，他可以根据新的ij非交换规则，将两个任意共面三元组相乘，从而进一步测试他的模型。尽管最初他很难定义新产品的方向，但经过多次尝试，他意识到毕达哥拉斯定理可以很好地解决这个问题。这又是一个令人鼓舞的成就。然而，这最后一步又给他带来了另一个问题：这个共面乘法的代数表示和几何表示相差 ( bz - cy ) ²倍。“我必须找到一种方法来删除这个多余的项，”他想。“我不想把事情从头开始！”

Dublin, fall of 1843. William Rowan Hamilton is in a challenging mood: even though he bumps into another impasse in his attempt to extend complex number theory to a three-dimensional space, he is obviously making important progress.²¹ He is particularly proud of his new starting point; what a mistake it was to start his previous experiments from tiring algebraic models! As he now starts geometrically by moving from x + iy to x + iy + jz, he possesses a three-dimensional line segment that is far easier to test (even though it adds a second imaginary number j right from the start). His first experiment was, in that sense, very conclusive. Thanks to the advice of his German colleague Gotthold Eisenstein, he could reach an equivalence between algebraic and geometrical definitions of the square of his three-dimensional segment by abandoning the assumption of commutation between i and j. He could then further test his model by multiplying two arbitrary coplanar triplets according to his new noncommutative rule for ij. Although he struggled at first to define the orientation of his new product, he realized—after several attempts—that Pythagoras’s theorem could nicely do the trick. Here again, an encouraging achievement. Yet this last move led him to another problem: the algebraic and geometrical representations of this coplanar multiplication differ by a factor of (bz—cy)². “I must find a way to remove this superfluous term,” he thinks. “I don’t want to start the whole thing over again!”

尽管这两个场景看起来有些神秘，但它们能告诉我们有关实验室实践的什么信息？我们能否在吉尔曼的内分泌实验室（场景 4）和汉密尔顿的数学实验室（场景 5）中发现相似之处？

Despite their cryptic aspects, what do these two scenes tell us about laboratory practices? Can we draw similarities between what takes place within Guillemin’s laboratory of endocrinology (scene 4) and what takes place within Hamilton’s laboratory of mathematics (scene 5)?

我们首先注意到，这两个场景都涉及实验；它们都对某物进行测试以评估其反应。场景 4 中的肽在 1973 年仍未定义。与最近关于这类氨基酸聚合物的说法一致，吉尔曼确信它应该会触发老鼠的生长激素。22^但这种生长激素被触发的程度如何？在什么情况下？为了更清楚地了解这种肽的能力，他让布拉索负责实施他最近设计的一项实验。在场景 5 中，复杂的三维线段x + iy + jz在 1843^年仍未定义。23汉密尔顿希望这个“三重态”——他称之为——能让他扩展复数论的几何表示。24^但目前还不确定。为了更好地了解他的复杂三维线段的能力，他对其进行了两次连续的实验：他首先将其平方，然后将它与另一个任意共面三重态相乘。

We can first notice that both scenes deal with experiments; they both put something to the test in order to evaluate its reactions. The peptide in scene 4 is, in 1973, still undefined. Guillemin—in line with recent claims about this class of amino acid polymer—is convinced that it should trigger the rat’s growth hormone.²² But how much is such growth hormone triggered? And under what circumstances? To have a clearer view on the capacities of this peptide, he puts Brazeau in charge of implementing an experiment he recently designed. In scene 5, a complex three-dimensional line segment x + iy + jz is, in 1843, still undefined.²³ Hamilton hopes that this “triplet”—as he calls it—will allow him to extend the geometrical representation of complex number theory.²⁴ But at this point, nothing is certain. To better understand the capacities of his complex three-dimensional line segment, he puts it through two successive experiments: he first squares it and then multiplies it with another arbitrary coplanar triplet.

在这两个场景中，实验都是为了测试未定义的实体而进行的。然而，实验不会自动发生；在这两个场景中，科学家都使用仪器来帮助他们探测未定义的实体。在场景 4 中，精心组装的大鼠垂体细胞培养物和非常广泛的放射免疫测定法是用于测试肽的两种主要工具。值得注意的是，这两种仪器都非常显眼，占据了很大的空间。场景 5 中的仪器先验上不那么令人印象深刻，但同样重要。第一个仪器显然是代数仪器正如中世纪伊斯兰数学家逐步定义的那样；如果没有任何手段以简练简洁的方式表达变量之间的关系，汉密尔顿就无法处理他的三重态。25^但他也需要一个坐标空间来几何地表达他的三重态。从这个意义上说，如果没有笛卡尔、费马、牛顿和莱布尼茨等十七世纪数学家的努力，汉密尔顿就没有办法考虑他的三重态的变换。他还需要一些非交换代数的见解，正如戈特霍尔德·艾森斯坦最近提出的那样，来处理复数乘积ij (Hankins 1980)。最后，他需要古老的毕达哥拉斯定理将他的初始三重态与另一个任意共面三重态相乘^。26

In both scenes then, experiments are run to test undefined entities. Yet experiments do not happen by themselves; in both scenes, instruments are used by scientists in order to help them probe their undefined entities. In scene 4, the delicately assembled rat pituitary cell culture and the very expansive radioimmunoassay are the two principal tools used to test the peptide. It is worth noting that both instruments are highly visible and take up a lot of space. The instruments in scene 5 are a priori less impressive but equally important. The first instrument is, obviously, the algebraic apparatus as progressively defined by medieval Islamic mathematicians; without any means to express relationships among variables in a condensed and succinct manner, Hamilton could not juggle his triplet.²⁵ But he also needs a coordinate space to express his triplet geometrically. In that sense, without the efforts of seventeenth-century mathematicians such as Descartes, de Fermat, Newton, and Leibniz, Hamilton would have no means to consider the transformations of his triplet. He further requires some insight from noncommutative algebra, as then recently proposed by Gotthold Eisenstein, to handle the complex product ij (Hankins 1980). Finally, he needs good old Pythagoras’s theorem to multiply his initial triplet with another arbitrary coplanar triplet.²⁶

此时，我们需要再进行一次务实的观察：尽管两个实验室都有对未定义实体进行实验的仪器，但这些仪器的形状各不相同。一方面，有一个生物电子组合，里面聚集了肽、布拉索、老鼠细胞、实验室技术人员和一个装满电子零件的巨型金属盒；另一方面，还有书籍、纸张、汉密尔顿和一支铅笔。这里几乎没有什么可怀疑的余地：这些仪器占用的空间并不相同。汉密尔顿的仪器看起来更干更薄，而吉尔曼的仪器看起来更湿更厚。有人可能会说——这也是我在本节剩余部分将使用的术语——汉密尔顿的实验室是平坦的，而吉尔曼的实验室是笨重的。两个实验室都从事相同的过程——测试未定义实体的反应——但它们使用的仪器在占用空间方面不同^。27

At this point, we need to make another down-to-earth observation: although both laboratories have instruments to conduct experiments on undefined entities, the shapes of these instruments differ from each other. On the one hand, there is a bioelectronic assemblage that gathers peptides, Brazeau, rat cells, laboratory technicians, and an imposing metal box full of electronic parts; on the other hand, there are books, paper, Hamilton, and a pencil. There is little room for doubt here: the instruments do not take up the same amount of space. Hamilton’s instruments appear dryer and thinner whereas Guillemin’s instruments appear wetter and thicker. One could say—and that is the terminology I will use for the remainder of this section—that Hamilton’s laboratory is flat whereas Guillemin’s laboratory is bulky. Both laboratories are engaged in the same process—testing the reactions of an undefined entity—but they use instruments that are different in terms of occupied space.²⁷

我们是否可以反过来说，吉尔曼的实验室比汉密尔顿的实验室更宽敞？如果我们只考虑他们仪器的相对价格，情况似乎确实如此：纸张比实验室技术人员便宜，大多数书籍（即使在十九世纪的爱尔兰）也比 20 世纪 70 年代的放射免疫测定便宜，铅笔比大鼠垂体细胞培养便宜。然而，如果考虑两种实验室设备的相对网络，这个问题似乎更棘手。事实上，培养和销售标准化大鼠细胞需要多少努力？毫无疑问，需要很多努力。但建立坐标空间需要多少努力？毫无疑问，需要很多努力。代数呢？正如 Netz (1998, 2004) 所表明的那样，没有对希腊几何著作的几个世纪评论，没有拜占庭图书馆，如果没有巴格达数学家的分类工作，任何代数符号系统都不可能存在。毕达哥拉斯定理也是如此；从古代早期到十九世纪的爱尔兰，收集、编纂和保存毕达哥拉斯命题需要许多长期的努力。让我们继续讨论我们两个实验室之间的拓扑差异：汉密尔顿的实验室比吉尔曼的实验室更平坦。

Can we in turn say that Guillemin’s laboratory is more expansive than Hamilton’s laboratory? If we only consider the relative price of their instruments, it seems indeed to be the case: paper is cheaper than laboratory technicians, most books (even in nineteenth-century Ireland) are cheaper than a radioimmunoassay from the 1970s, and pencils are cheaper than a rat pituitary cell culture. Yet if one considers the relative networks of both laboratory apparatuses, the question appears trickier. Indeed, how many efforts were needed to cultivate and sell standardized rat cells? Many, indubitably. But how many efforts were required to establish coordinate spaces? Many, indubitably. And what about algebra? As Netz (1998, 2004) showed, without centuries of commentaries on Greek geometrical writings, without Byzantine libraries, and without the classification efforts of Bagdad mathematicians, no algebraic system of notation could have come into existence. The same is true of Pythagoras’s theorem; many long-standing efforts were required to gather, compile, and preserve Pythagorean propositions from early antiquity to nineteenth-century Ireland. Let us then stick to the topological difference between our two laboratories: Hamilton’s laboratory is flatter than Guillemin’s.

如果我们继续分析这两个场景，我们可以看到，尽管它们在拓扑结构上存在差异，但笨重和扁平的仪器最终都会产生类似的铭文；即文件上的可读痕迹。事实上，场景 4 中笨重的生物电子实验组合最终产生的曲线表明老鼠的激素减少了。布拉索对未定义肽进行的实验结果只是吉尔曼焦急地检查的纸片。28^同样，场景 5 中的扁平实验组合最终产生了一系列耦合的代数和几何方程；起初，两个方程看起来是等价的（这对汉密尔顿来说是个好消息），但在实验的第二步，两者看起来并不相似（这对汉密尔顿来说是个坏消息）。然而，就像布拉索和吉尔曼一样，汉密尔顿的扁平实验的结果就是他用眼睛检查的文件上的可读痕迹^。29

If we continue to analyze both scenes, we can see that despite their topological differences, both bulky and flat instruments end up producing comparable inscriptions; that is, readable traces on documents. Indeed, the bulky bioelectronic experimental assemblage of scene 4 ends up producing graphs whose curves indicate that the rat’s hormone decreases. The results of the experiment on the undefined peptide conducted by Brazeau are pieces of paper anxiously examined by Guillemin.²⁸ Similarly, the flat experimental assemblage of scene 5 ends up producing a series of coupled algebraic and geometrical equations; at first, both equations appeared equivalent (which was good news for Hamilton), but in the second step of the experiment, both appeared dissimilar (which was bad news for Hamilton). Yet, just as for Brazeau and Guillemin, the results of Hamilton’s flat experiments are readable traces on documents he examines with his eyes.²⁹

至此，我们可以暂时说这两个场景都涉及实验、仪器（不同拓扑结构）和一系列铭文。但所有这些工作会导向何处？目前，它当然不会导致任何可能后来成为科学事实的已发表声明。在这两个实验室中，科学家对未定义的实体进行测试，但这些实践如何导致能够在学术论文中描述的对象的形成？

At this point then, we can tentatively say that both scenes deal with experiments, instruments (of different topologies), and series of inscriptions. But where does all this work lead to? At this stage, it certainly cannot lead to any published claim that may later become a scientific fact. Within these two laboratories, scientists impose tests on undefined entities, but how can these practices lead to the formation of objects capable of being described in academic papers?

场景 6

Scene 6

索尔克生物研究所，拉霍亚（加利福尼亚州），1973 年 1 月^。30对此无能为力；即使在另外两次细致的实验之后，放射免疫分析打印的图表仍然显示，当与 Guillemin 的肽接触时，大鼠的激素会减少。大鼠垂体细胞培养物是无可争议的，Guillemin 的肽的成分、放射免疫分析和 Brazeau 的专业精神也是如此（Guillemin 很快就承认了这一点）。摆脱这种僵局的唯一方法就是对这种肽的作用提出质疑。包括 Guillemin 在内的内分泌学领军人物认为这类肽会触发生长激素；显然，它的作用恰恰相反。在与大鼠垂体细胞培养物接触一段时间后，经过一些一致参数的放射免疫测定后，这种新东西会显著降低大鼠的生长激素。由于可以肯定实验过程中没有出现任何错误，现在正在准备一篇论文，以说服持怀疑态度的读者相信这种新的科学对象的存在，Guillemin 开始将其称为生长抑素（字面意思是“阻断身体的东西”）。

Salk Institute for Biological Studies at La Jolla (California), January 1973.³⁰ There is nothing to do about it; even after two other meticulous experiments, the graphs printed by the radioimmunoassay still show that the rat’s hormone decreases when put in contact with Guillemin’s peptide. The rat pituitary cell culture is indisputable as are the composition of Guillemin’s peptide, the radioimmunoassay, and Brazeau’s professionalism (Guillemin quickly admits it). The only way to escape from this impasse is to cast doubt on what the peptide does. Leading figures in endocrinology—including Guillemin—thought that this class of peptide triggered the growth hormone; obviously, it does the opposite. After being in contact with rat pituitary cell culture for a certain amount of time and after having gone through the radioimmunoassay with some consistent parameters, this new thing significantly decreases the rat’s growth hormone. As it is certain that there have been no mistakes during the experimental procedures, a paper is now being prepared to convince skeptical readers about the existence of this new scientific object Guillemin starts to call somatostatin (literally, “that which blocks the body”).

场景 7

Scene 7

都柏林，1843 年秋^。31对此无能为力：如果不添加新的虚量，就无法删除复线段长度的几何表达式中的多余项 ( bz — cy ) ²。代数规则（包括非交换性）是不容置疑的，毕达哥拉斯定理和汉密尔顿的经典操作（他多次进行了整个实验）也是如此。摆脱这种僵局的唯一方法是质疑实验的前提：如果复数论的几何表示的扩展需要四个维度而不是三个维度会怎样？事实上，只有加入第三个虚量k作为i和j的乘积才能使多余的项 ( bz — cy ) ²消失。诚然，这个新的虚量需要第四个轴才能进行几何表示，但是谁在乎呢？在引入k作为虚数（代数表示）或四维轴（几何表示）后，这个新事物可以平方和相乘，同时产生等效方程，从而有效地扩展复数理论的几何表示。如果汉密尔顿现在设法定义k 2、ik、kj和i 2这些量（目前几乎只是走个过场），他将能够完全定义这个他开始称之为四元数（字面意思是“由四个组成的事物”）的新数学对象的行为。

Dublin, fall of 1843.³¹ There is nothing to do about it: the superfluous term (bz—cy)² within the geometrical expression of the length of a complex line segment cannot be removed without adding a new imaginary quantity. The rules of algebra—including noncommutativity—are indisputable, as are Pythagoras’s theorem and Hamilton’s scriptural operations (he ran the whole experiment several times). The only way to escape from this impasse is to cast doubt on the premises of the experiment: What if the extension of the geometrical representation of complex number theory required not three but four dimensions? Indeed, only the inclusion of a third imaginary quantity k as the product of i and j can make the superfluous term (bz—cy)² disappear. It is true that this new imaginary quantity needs in turn a fourth axis in order to be geometrically represented, but who cares? After the introduction of k as either an imaginary quantity (in the algebraic representation) or a fourth dimensional axis (in the geometrical representation), this new thing can be squared and multiplied while producing equivalent equations, hence effectively extending the geometrical representation of complex number theory. If Hamilton now manages to define the quantities k2, ik, kj, and i2—almost a formality at this stage—he will be able to completely define the behavior of this new mathematical object he starts to call quaternion (literally, “that which is made of four”).

那么，除了神秘的一面之外，这两个场景还告诉我们什么有关科学实验室内新物体形成的信息呢？我们能否得出生长抑素（场景 6）和四元数（场景 7）的渐进成型之间是否存在相似之处？

Again, beyond their cryptic aspects, what do these two scenes tell us about the formation of new objects within scientific laboratories? Can we draw some similarities between the progressive shaping of somatostatin (scene 6) and quaternions (scene 7)?

我们首先可以看到，在两个场景中，仪器打印出来的文字都是从表达单一现象开始的。在场景 6 中，放射免疫分析打印出来的图表自信地表明，在将肽注入大鼠垂体细胞培养物中一段时间​​，并经过特定参数的放射免疫分析后，生长激素会显著减少。这正是 Guillemin 可以读懂的图表中所写的内容；整个实验过程最终导致大鼠的生长激素减少。可信的图表变得更平坦；因此生长激素会减少。

We can first see that in both scenes, inscriptions printed out by instruments begin by expressing singular phenomena. In scene 6, the graphs printed by the radioimmunoassay indicate confidently that after the peptide is injected in the rat pituitary cell culture over a specific period of time and after it goes through the radioimmunoassay with specific parameters, the growth hormone decreases significantly. This is what is inscribed within the graphs Guillemin can read; the whole experimental process ends up decreasing the rat’s growth hormone. Trustful graphs become flatter; therefore the growth hormone decreases.

同样，在场景 7 中，汉密尔顿手中刻下的铭文表明，在将第四维添加到三元组之后，为了用几何方式表达新的虚数k（它本身是使多余项 ( bz — cy ) ²消失所必需的），复数论的代数表示和几何表示都变得等价。这也是汉密尔顿在纸上读到的铭文所描述的现象；整个实验过程最终表达了复数论的几何表示和代数表示之间等价性的扩展。一个可信的几何方程变得等价于另一个代数方程；因此，复数论的几何表示得到了扩展。

Similarly, in scene 7, the inscriptions produced by the hands of Hamilton indicate that after a fourth dimension is added to the triplet in order to geometrically express the new imaginary quantity k—itself required to make the superfluous term (bz—cy)² disappear—both algebraic and geometrical representations of complex number theory become equivalent. Again, this is the phenomenon described by the inscriptions Hamilton can read on a sheet of paper; the whole experimental process ends up expressing an extension of the equivalence between geometrical and algebraic representation of complex number theory. A trustful geometrical equation becomes equivalent to another algebraic equation; therefore, the geometrical representation of complex number theory is extended.

然而，这也是关键所在，凭借实验设置，这两个现象的起源——“生长激素的可量化抑制”和“几何与复数理论之间的等价性的扩展”——可以归因于具体的事物。在场景 6 中，实验过程开始时唯一未定义其行为的元素是肽。大鼠垂体细胞培养、放射免疫分析、布拉索和技术人员的行为都是可以预测的；因此，不可预测的现象——图形变得更平坦——一定是来自这种“阻塞身体”的肽类东西的行为。同样，在场景 7 中，实验设置的这个阶段唯一未定义其行为的元素是第三虚数k，它通过第四维轴以几何形式表示。非交换代数、毕达哥拉斯定理和汉密尔顿的纸上操作的行为都是可以预测的；几何方程和代数方程变得等价这一不可预测但又在意料之中的现象只能是归因于这个“将四个数字组合在一起”的四维事物。在这两个场景中，新事物都来自相同的归因过程；新现象的经文痕迹被归因于先前未定义的实体的行为。

However, and this is the crucial point, by virtue of the experimental setting, the origins of these two phenomena—“quantifiable inhibition of the growth hormone” and “extension of the equivalence between geometry and complex number theory”—can be attributed to specific things. In scene 6, the only element whose actions were undefined at the beginning of the experimental process was the peptide. The actions of rat pituitary cell cultures, radioimmunoassay, Brazeau, and the technicians were all predictable; the unpredictable phenomenon—the graphs becoming flatter—must thus result from the action of this peptide-thing that “blocks the body.” Similarly, in scene 7, the only element whose actions were undefined at this stage of the experimental setting was the third imaginary quantity k geometrically expressed by a fourth dimensional axis. The actions of noncommutative algebra, Pythagoras’s theorem, and Hamilton’s pencil and paper operations were all predictable; the unpredictable, yet anticipated, phenomenon—geometrical and algebraic equations becoming equivalent—can only be attributed to this four-dimensional thing that “groups together four numbers.” In both scenes, new things emerge from the same attribution process; scriptural traces of a new phenomenon are imputed to the behavior of a previously undefined entity.

在两个场景的结尾，这种根据实验设置将某种行为归因于先前未定义的实体的归因过程最终被一个术语所概括，这个术语概括了现已定义的事物的作用：“阻碍身体的事物”变成了生长抑素，“将四个数字组合在一起的事物”变成了四元数。新的事物出现了，但并没有什么奇迹。在这两种情况下，新事物的形状都是随着科学家让它从一系列动作“成长”为事物的名称而逐渐定义的。在场景 6 中，生长抑素首先是“图形变得更平坦”，然后是“在这些实验条件下，生长激素减少”，然后是“我们的新肽降低了老鼠的生长激素”，最后是“生长抑素降低了老鼠的生长激素”。同样的具体化过程（Latour 1987, 86–100）也发生在场景 7 中：四元数首先是“两个方程变得等价”，然后是“复数论的几何表示的扩展”，然后是“四维表示允许复数论的几何表示的扩展”，最后是“四元数在四维空间中以几何方式表达复数论”。在这两种情况下，实验、仪器和铭文的排列——简而言之，实验室实践（Latour and Woolgar 1986）——逐渐导致了科学对象的形成，而这些对象的属性和轮廓反过来又可以成为声称其存在的论文的主题^。32

At the end of both scenes, this attribution process that imputes a behavior to a previously undefined entity by virtue of an experimental setting ends up being summarized by a term that encapsulates what the now defined thing does: “that which blocks the body” becomes somatostatin and “that which groups four numbers” becomes quaternion. New objects come into existence, but there has been no miracle; in both cases, the shape of the new object was progressively defined as scientists made it “grow” from a list of actions to the name of a thing. In scene 6, somatostatin was first “the graphs become flatter,” then “under these experimental conditions, there is a diminution of the growth hormone,” then “our new peptide decreases rat’s growth hormone,” and finally “somatostatin decreases rat’s growth hormone.” The same reification process (Latour 1987, 86–100) happened in scene 7: quaternion was first “two equations become equivalent,” then “there is an extension of geometrical representation of complex number theory,” then “four-dimensional representation allows the extension of geometrical representation of complex number theory,” and finally “quaternions express geometrically complex number theory in a four-dimensional space.” In both cases, experiments, instruments, and alignments of inscriptions—in short, laboratory practices (Latour and Woolgar 1986)—progressively led to the shaping of scientific objects whose properties and contours could, in turn, become the topics of papers claiming their existence.³²

然而，正如我们在上一节中看到的，无论是生长抑素还是四元数，在论文中呈现给持怀疑态度的同事，都需要克服许多考验才能成为经过认证的科学事实，能够被黑箱化、风格化、完善，并纳入进一步的主张和实验环境。尽管这两个对象都是在各自庞大而平坦的实验室中诞生的，但它们仍然需要吸引更广泛社区的支持。但是，当怀疑读者的疑虑消除后，当科学机构认证了这两个说法的真实性后，我们就可以自信地说，吉尔曼发现了生长抑素，汉密尔顿发现了四元数。或者我们可以吗？我们确实看到，这两个对象都是实验室实践的结果，这些实践逐渐塑造了它们。科学家能发现他们发现的对象吗？之前构建的科学对象？生长抑素和四元数是否已经是“自然”的一部分，即使它们必须在设备齐全（但拓扑结构不同）的实验室中形成？这就是故事开始变得棘手的地方。如果STS早已表明科学对象需要在实验室中制造，那么一旦关于科学对象的书面声明成为经过认证的事实，这些地点的重型设备以及使它们发挥作用所需的实际工作就会消失。一旦不再有关于新科学对象的争议或分歧，自然界往往会被视为始终包含这个已构建的科学对象的领域。在这里，我们遇到了我们在第4章讨论计算机编程实践时讨论过的事情：当事实得到认证并纳入进一步研究时，允许它们逐步形成的实验、仪器、社区和实践通常会被搁置一旁（Latour and Woolgar 1986，105-155）。这就是科学史和社会学（包括数学）如此难以进行的原因；当既定事实从支持其形成的人为背景中提炼出来时，人们很容易从这些既定事实出发，向后推断（Collins 1975）^。33

However, as we saw in the previous section, both somatostatin and quaternions as presented in papers that can be read by skeptical colleagues still need to overcome many trials to become certified scientific facts capable of being blackboxed, stylized, polished, and enrolled in further claims and experimental settings. Although both objects came into existence within their respective bulky and flat laboratories, they still need to attract the adherence of a wider community. But when the doubts of skeptical readers are removed, when the veracity of both claims are certified by the scientific institution, we can in turn confidently say that Guillemin discovered somatostatin and that Hamilton discovered quaternions. Or can we? We saw indeed that both objects were the results of laboratory practices that progressively shaped them. Can scientists discover objects they were previously constructing? Were somatostatin and quaternions already part of “nature” even though they had to be shaped in well-equipped (yet topologically different) laboratories? This is where the story starts to become tricky. If STS has long shown that scientific objects need to be manufactured in laboratories, the heavy apparatus of these locations as well as the practical work needed to make them operative tend to vanish as soon as written claims about scientific objects become certified facts. Once there are no more controversies or disagreements about a new scientific object, nature tends to be invoked as the realm that always already contained this constructed scientific object. Here, we encounter something we discussed in chapter 4 where we were dealing with computer programming practices: when facts are certified and enrolled in further studies, the experiments, instruments, communities, and practices that allowed their progressive formation are generally put aside (Latour and Woolgar 1986, 105–155). This is what makes the history and sociology of sciences (including mathematics) so difficult to conduct; as established facts are purified from the artificial setting that supported their formation, the temptation is great to start from these established facts and extrapolate backward (Collins 1975).³³

然而，如果一个人对科学史或科学社会学不感兴趣，如果一个人“只是”想谈论客观事实并最终将它们纳入进一步的主张，那么提及自然似乎是完全合理的。从这个意义上说，当然可以说——作为一种方便的捷径——汉密尔顿“发现”了四元数，或者吉尔曼“发现”了生长抑素，但这只是因为这些对象最终被接受为经过认证的事实，被放入黑匣子中，翻译、润色，并纳入后来的主张中。由于最初在书面声明中提出的制造对象相继抵制了试验，因此它们在专用实验室中的生产条件可以暂时被忽视；自然可以接管并支持它们的存在理由。在这方面，拉图尔的有趣类比很有启发性：

However, if one is not interested in the history or sociology of sciences, if one “just” wants to speak about objective facts and eventually enroll them in further claims, the reference to nature appears completely justified. In that sense, one may of course say—as a kind of convenient shortcut—that Hamilton “discovered” quaternions or that Guillemin “discovered” somatostatin, but only because these objects ended up being accepted as certified facts, put in black boxes, translated, polished, and enrolled in later claims. As both initially manufactured objects presented in written claims successively resisted trials, the conditions of their production within dedicated laboratories can be, temporarily, neglected; nature can take over and support their raison d’être. In this respect, Latour’s funny analogy is quite instructive:

在科学家的手中，自然是一位立宪君主，就像伊丽莎白二世女王一样。她坐在宝座上，用同样的语气、威严和信念宣读保守党或工党首相根据选举结果撰写的演讲稿。事实上，她为争议增添了一些内容，但只是在争议结束后；只要选举还在进行，她就什么也不做，只是等待。（Latour 1987，98）

Nature, in scientists’ hands, is a constitutional monarch, much like Queen Elizabeth the Second. From the throne she reads with the same tone, majesty and conviction, a speech written by Conservative or Labour prime ministers depending on the election outcome. Indeed she adds something to the dispute, but only after the dispute has ended; as long as the election is going on she does nothing but wait. (Latour 1987, 98)

因此，“自然”这个概念对于谈论无争议的科学事实很方便——为什么不呢？——但一旦人们谈论科学争论或关于科学对象的形成，人们需要将自然视为科学实践的不确定结果。34这种对自然的谨慎态度^既适用于“传统的”笨重科学对象，如生长抑素，也适用于“非常规的”扁平科学对象，如四元数。同样，没有哪个优越的现实能让数学对象出现在数学家面前。它们也需要在配备打印铭文的仪器的（扁平）实验室内成形。

The notion of “nature” is thus convenient to speak about noncontroversial scientific facts—why not?—but as soon as one speaks about scientific controversies or about scientific objects in the making, one needs to consider nature as the uncertain result of scientific practices.³⁴ This cautious position toward nature applies to “conventional” bulky scientific objects such as somatostatin as well as to “unconventional” flat scientific objects such as quaternions. Again, no superior reality makes mathematical objects appear to mathematicians. They too need to be shaped within (flat) laboratories equipped with instruments that print inscriptions.

数学有能力的

Mathematicable

一件好事已经得到解决：科学事实的构建过程似乎确实与数学事实的构建过程非常相似。定理（参见场景 1 和 3）、数学系统（参见场景 5 和 7）、猜想（参见场景 2）甚至公式（参见场景 3）都可以被视为真正的科学主张，试图说服同事相信先前在（平面）实验室内形成的物体的存在。如果这些主张中的绝大多数无法通过考验成为经过认证的事实，其中一些（例如，香农-哈特利定理、汉密尔顿的四元数理论）可能会成为程式化和精致的黑匣子，用作进一步实验环境的工具。我们可以将这个庞大且不断变化的经过认证的数学事实库称为“数学知识”。此外，这个经过认证的知识体系的几个元素有时可能会成为隐性、无可争辩和必要的知识的一部分（例如，逻辑推理定律）。

A good thing has been taken care of: it seems indeed that the construction process of scientific facts is quite similar to the construction process of mathematical facts. Theorems (cf. scenes 1 and 3), mathematical systems (cf. scenes 5 and 7), conjectures (cf. scene 2), and even formulas (cf. scene 3) may all be considered genuine scientific claims that try to convince colleagues of the existence of objects previously shaped within (flat) laboratories. If the vast majority of these claims do not overcome the trials that can make them become certified facts, some of them (e.g., Shannon-Hartley’s theorem, Hamilton’s theory of quaternions) may become stylized and polished black boxes that are used as instruments in further experimental settings. It is this huge—and changing—repository of certified mathematical facts that we may call “mathematical knowledge.” Moreover, several elements of this certified body of knowledge may, sometimes, become part of tacit, indisputable, and necessary knowledge (e.g., the logical laws of deduction).

然而，尽管各自的构建过程有着惊人的相似之处，已证实的科学和数学事实及其相关对象似乎仍然存在显著差异：

However, despite the striking similarities between their respective construction processes, certified scientific and mathematical facts—and their correlated objects—still seem to differ significantly:

持怀疑态度的读者的反对意见

Objection of a skeptical reader

好吧，让我们假设事实和相关对象都经历了类似的构造过程，正如你显然相信的那样（同时仅依赖于小的、不完整的示例）。一个重要的区别是：数学对象永远不会停止用于构造非数学对象！我们甚至可以在你用来说明你的观点的内分泌实验室中看到它。放射免疫分析打印的图表量化了生长激素的含量肽减少的输入是固化的数学事实（在本例中是基本解析几何）。放射免疫测定的内部机制当然也是如此；必须使用复杂的数学理论来开发这种昂贵的仪器。人口统计学、气候学、政治学、生物学等领域一直在发生类似的过程。对数、高斯函数或概率等数学对象渗透到“硬”科学的所有领域，帮助科学家塑造新的对象和事实。然而，反之则不然：肽或放射免疫测定如何帮助数学家塑造新对象？数学家必须自己做事，而不需要其他科学的帮助。这就是为什么数学是所有科学中的王后：如果没有数学家在“平坦实验室”中的工作——我们可以这样说——就不会有精确的科学。数学对象如此强大；它们一定具有某种优越的性质。怎么可能不是呢？

All right, let’s assume that both facts—and correlated objects—go through similar construction processes, as you obviously believe (while only relying on small, incomplete examples). An important difference subsists: mathematical objects never stop being used for the constitution of nonmathematical objects! We could even see it in the laboratory of endocrinology you used to illustrate your point. The graphs printed by the radioimmunoassay, which quantify how much the growth hormone is decreased by the peptide, are importations of solidified mathematical facts (in this case, basic analytical geometry). The same is certainly true of the inner mechanisms of the radioimmunoassay; complex mathematical theories must have been used to develop this costly instrument. Similar processes happen all the time in demography, climatology, political science, biology, and so on. Mathematical objects such as logarithms, Gaussian functions, or probabilities infiltrate all domains of “hard” science, helping scientists to shape new objects and facts. Yet the inverse is not true: how could peptides or radioimmunoassay help mathematicians shape new objects? Mathematicians have to do things by themselves, without the help of the other sciences. This is why mathematics is the queen of all sciences: without the work of mathematicians in their “flat laboratories”—we may keep that—there would simply be no exact sciences. Mathematical objects are so powerful; they must be of some superior nature. How could it be otherwise?

这种经典的反对意见有两个缺陷。首先，数学实践是自给自足的说法站不住脚，因为许多学科都干预了数学对象和事实的构建过程。例如，Netz (1998, 2004) 展示了存档和标准化如何成为克服希腊几何学停滞的关键。35^由于收集了精心整理的纸莎草纸和羊皮纸资料（尤其是在拜占庭），像 Eutocius 这样的晚期古代评论家能够比较、注释和完成希腊几何著作中错综复杂的多样性。逐渐地，这些系统的标准化努​​力使得早期古代的几何命题具有可比性；与希腊几何学家不同，³⁶中世纪数学家（尤其是在巴格达的智慧之家）（Netz 2004, 131–186）能够看到希腊几何学是什么。凭借“智力技术”（Goody 1977）——这里指的是标准化的希腊几何论文集——数学家（如花拉子米和 Khayyam）能够系统化和分类希腊人解决的几何问题。根据 Netz 的说法，这些系统化的比较逐渐导致了代数语言的形成：“花拉子米的代数最终是一个相当不雄心勃勃的野心，转化为重大变革。花拉子米除了对过去的结果进行分类之外，并没有做任何其他事情，他实际上创造了方程式”（Netz 2004，143）。

There are two glitches in this classical objection. First, it is not tenable to say that the practice of mathematics is self-sufficient, for many disciplines intervene in the construction process of mathematical objects and facts. Netz (1998, 2004) showed, for example, how archiving and standardization were central to overcome the stagnation of Greek geometry.³⁵ Thanks to the assembling of well-arrayed corpora of papyruses and parchments—especially in Byzantium—late antiquity commentators such as Eutocius became able to compare, annotate, and complete the entangled multiplicities of Greek geometrical writings. Progressively, these systematic standardization efforts made early antiquity’s geometrical propositions commensurable; unlike Greek geometers,³⁶ medieval mathematicians—especially in Bagdad’s House of Wisdom (Netz 2004, 131–186)—could see what Greek geometry was. Equipped with “intellectual technologies” (Goody 1977)—here, collections of standardized Greek geometrical treatises—mathematicians such as al-Khwarizmi and Khayyam could systematize and classify the geometrical problems solved by the Greeks. These systematic comparisons progressively led, according to Netz, to the formation of the algebraic language: “Al-Khwarizmi’s algebra was, ultimately, a fairly unambitious ambition, translated into major transformations. Without himself doing anything beyond classifying the results of the past, Al-Khwarizmi, effectively, created the equation” (Netz 2004, 143).

既然归档和标准化曾经是、现在也是^数学对象形成的核心，我们是否必须说这两个受人尊敬的学科是所有科学中的王后？对我来说，更合理的观点是承认学科的等级分类具有误导性。当某事物允许其他事物存在时，它可能不是垂直等级的问题，而是水平排列的问题。

Since archiving and standardization were, and are,³⁷ central to the formation of mathematical objects, do we have to say that these two respectable disciplines are the queens of the queen of all sciences? To me, a more reasonable position would be to accept that hierarchal classification of disciplines is misleading. When something allows something else to come into existence, it may not be a matter of vertical hierarchy but of horizontal arrangement.

这引出了第二个反对意见，即数学对象是否可用于组装非数学对象。数学事实的组合能力确实令人惊讶。在每一门科学学科中，近期或远古的数学发现都用于进行实验、整理铭文、表达新现象，并最终定义新对象。我甚至会比我们的怀疑论读者走得更远，将数学对象的这种极端组合性扩展到日常生活中。例如，我们一天使用算术的基本规则多少次？显然，数学无处不在，从高能物理实验室到收银台。这种渗透到异质活动领域的能力非常令人印象深刻。但这是否必然意味着数学对象来自不同的性质？它们的可塑性是否必然体现出超自然的本质？

This leads us to the second objection regarding the usability of mathematical objects for the assembling of nonmathematical objects. It is true that the combinational capabilities of mathematical facts are surprising. In every scientific discipline, recent or ancient mathematical discoveries are used to conduct experiments, organize inscriptions, express new phenomena, and eventually define new objects. I would go even further than our skeptical reader and expand this extreme combinability of mathematical objects to everyday life. For example, how many times a day do we use the basic precepts of arithmetic? Obviously, mathematics is everywhere, from laboratories of high energy physics to cashiers’ desks. This capacity to infiltrate heterogeneous domains of activity is very impressive. But does it necessarily mean that mathematical objects come from a different nature? Does their plasticity necessarily manifest a supernatural essence?

让我们以 Guillemin 的内分泌实验室为例，因为这是我们持怀疑态度的读者使用的例子。的确，放射免疫分析计算机打印的结果需要应用初等数学理论才能表明生长激素的减少。这其中有什么魔力吗？如果我们更精确地考虑将大鼠垂体细胞培养物“压平”为可表示为数值随时间变化的图形的过程，就不会有这种魔力了。放射免疫分析中究竟发生了什么？从示意图上看，大鼠垂体细胞培养物发射的非常微小的放射性波被捕获，经过一系列转换后，由昂贵的设备进行计数。放射性波变成信号，而信号又变成随时间变化的离散值。这种转化过程——或者更简洁地说，翻译过程——使细胞培养物从复杂液体状态转变为随时间推移的可写入（放射性）值列表状态，正是这种过程使得“比率”这一基本数学概念得以应用，并进一步计算生长激素的减少。毕达哥拉斯学派发展起来的祖先比率理论是如何适用于内分泌学领域？将细胞培养物以不同的方式（转化）为可量化的铭文，从而使其成为几何图形的具体努力，使比率与 Guillemin 的肽之间建立了联系。正是通过扁平化细胞培养物并使其适应扁平比率生态，这些数学对象才适用于细胞培养物。没有什么神秘的事情发生；通过逐步将复杂实体转化为圣经形式，可以将其与经过认证的数学事实联系起来。

Let us consider Guillemin’s laboratory of endocrinology since it is the example used by our skeptical reader. It is true that the results printed by the computer of the radioimmunoassay required the application of elementary mathematical theories in order to indicate a diminution of the growth hormone. Was there some magic? Not if we consider more precisely the process by which the rat pituitary cell culture was “flattened” to become representable as a graph with numerical values varying through time. What happened indeed within the radioimmunoassay? Schematically, the very small radioactive waves emitted by the rat pituitary cell culture were captured and, after a series of translations, counted by the costly equipment. Radioactive waves became signals that, in turn, became discrete values varying through time. This transubstantiation process—or, more succinctly, translation process—that made a cell culture go from the state of complex liquid to the state of a writable list of (radioactive) values spread over time is precisely what allowed the enrollment of the elementary mathematical notion of “ratio” and the further calculation of the growth hormone’s decreasing. How did the ancestral theory of ratios as developed by the Pythagoreans become applicable to the world of endocrinology? The concrete efforts to form differently (trans-form) the cell culture into quantifiable inscriptions, thus making it become a geometrical graph, allowed the connection between ratios and Guillemin’s peptide. It was by flattening the cell culture and adapting it to the flat ecology of ratios that these mathematical objects became applicable to the cell culture. Nothing mysterious happened; by progressively translating a complex entity into a scriptural form, it became possible to link it with certified mathematical facts.

另一个更好的例子是，这种使非数学实体变得可数学化的经验过程由 Michal Lynch (1985) 在其著作《实验室科学中的艺术与人工制品》中提供。20 世纪 70 年代，神经病学中的一个重要课题是大脑的可塑性；简而言之，就是大脑通过重组部分组织恢复失去的功能的能力。在 Lynch 进行实验室研究时，这种重组如何发生是一个有争议的话题。当时有两种主要猜想在相互竞争。第一种猜想认为重组是通过受损大脑区域内突触（允许轴突和树突之间进行神经元间通讯的结构）的致密化发生的^。38第二种理论被称为“轴突发芽”，认为重组是由于受损区域附近轴突的延伸造成的。由于多种原因（包括当时最近的实验室实验结果以及有希望的工业应用），Lynch 研究的实验室主任认为轴突发芽是大脑重组能力的主要因素（Lynch 1985, 32–33）。但他如何证明这一点呢？他遇到了许多陷阱。首先，神经元非常小。观察它们的（重新）组织需要强大的变焦能力。幸运的是，电子显微镜（实验室最近购买的一项技术）的出现使他能够进行超微结构观察。但这导致了另一个问题：当时，这些观察只能在微小的载玻片上进行，其平面拓扑结构与神经元的庞大拓扑结构不同。幸运的是，可以组织“一系列有条不紊的实验室大鼠渲染图”（Lynch 1985, 37）以正确切片大脑并使其适应超微结构可见性。但这种脑切片提取方法又引发了另一个问题，因为重组脑的过程只能在活脑中发生。那么，如何才能在死脑切片样本上观察到脑的可塑性呢？幸运的是，目前有许多标准化实验室大鼠，它们的大脑结构几乎完全相同。大脑允许组织“牺牲链”（Lynch 1985, 38）。虽然无法观察到一个活着的受损大脑的重组，但逐渐可以观察到在不同时间间隔杀死的“相同”受损大脑的重组。一系列有规律的离散的——并且经过精心参考的——死切片允许重建一个试图减轻其损伤的活大脑的进化。然而，Lynch 之后的科学家仍然需要在每张幻灯片的混乱中辨别出具体事件。他们确实试图解释轴突纤维将其领土扩展到损伤区。但他们如何定义轴突的区域以及它们的潜在扩张？幸运的是——这对整个项目的设计有很大帮助——“背海马”的一个有趣特征帮助他们建立了所有电子显微镜可观察切片的共同参考点。确实，已经证实并接受，背海马体的结构看起来像一个网格，其细胞体的树突有规律地与指向不同大脑区域的轴突相交（Ram ón y Cajal 1968）。因此，如果大脑研究人员设法从同样受损的老鼠大脑（在不同时间间隔杀死）中提取出电子显微镜可观察的背海马体切片，背海马体细胞的树突与指向不同大脑区域的轴突相交产生的“自然”网格结构可以构成进一步测量的初始经验基础（Lynch 1985，35-39）。换句话说，由于已经证明背海马的某个特定部分包含细胞体，而这些细胞体的树突总是与指向两个不同脑区（我在此称之为α和β ）的轴突有规律地相交，因此可以损伤所有大鼠的β脑区，然后检查指向α的轴突是否“发芽”并渗入先前指向β的轴突的区域。但这又带来了一个新问题：如何从切片上的特定电子显微镜视图转换为随时间分布的多个切片的全景图？在林奇进行研究时，操作这种转换的最简单方法是首先拍摄电子显微镜背海马显示的模拟照片。然后，脑科学家必须以高清晰度冲洗这些照片，并根据观察的超微结构水平（根据照片不同，缩放倍数在2,160到24,000倍之间）为它们配备一个坐标系。林奇的科学家们究竟是如何拍摄这些高清照片的？他们确定了这些照片被放在一张纸板上，从而创建了微观展示的按时间顺序排列的蒙太奇。正如林奇所说，“这些连续的照片提供了大脑超微结构的可见配置，这将在研究的分析阶段得到解决”（Lynch 1985, 38）。但在这里，仅仅测量以α为索引的轴突的延伸是不够的。即使背海马细胞体的树突有规律地与以α和β为索引的轴突相交，仍然需要为所有照片附加一个共同的参考。脑科学家们是如何做到这一点的？这里很难不引用林奇的说法：

Another—better—example of such an empirical process that makes non-mathematical entities become mathematicable is provided by Michal Lynch (1985) in his book Art and Artifact in Laboratory Science. During the 1970s, an important topic in neurology was the plasticity of the brain; that is—briefly stated—its capacity to recover lost functions through the reorganization of some of its tissues. How this reorganization occurs was a controversial topic at the time of Lynch’s laboratory study. Two major conjectures were in competition. The first one considered that the reorganization occurred through the densification of the synapses—the structures that allow interneuronal communication between axons and dendrites—within the damaged brain territory.³⁸ The second theory, labeled “axon sprouting,” considered that the reorganization was due to the extension of axons adjacent to the damaged territory. For many reasons encompassing results of then recent laboratory experiments as well as promising industrial applications, the director of the laboratory studied by Lynch believed that axon sprouting was the main ingredient for the brain’s reorganizational capacity (Lynch 1985, 32–33). But how could he demonstrate it? Many pitfalls got in his way. First, neurons are very small. Observing their (re)organization required powerful zooms. Fortunately, the advent of electron microscopy—a technology recently purchased by the laboratory—allowed him to make ultrastructural observations. But this led to another issue: at that time, these observations could only be made on tiny slides whose flat topology was different from the bulky topology of neurons. Fortunately, a “methodic series of renderings of laboratory rats” (Lynch 1985, 37) could be organized to properly slice brains and adapt them to ultrastructural visibility. But this extraction of brain slides led to another issue as a reorganizational brain process can only happen within a living brain. How could it then be possible to observe brain plasticity on dead sliced samples? Fortunately, the availability of many standardized laboratory rats with almost identical brains allowed the organization of a “chain of sacrifices” (Lynch 1985, 38). Although it was not possible to observe the reorganization of one living damaged brain, it progressively became possible to observe the reorganization of “same” damaged brains killed at different time intervals. A regular series of discrete—and meticulously referenced—dead slices permitted the reconstitution of the evolution of one living brain trying to palliate its damages. Yet the scientists followed by Lynch still needed to discern specific events within the mess of every single slide. They were indeed trying to account for axon fibers that were expanding their territories to damage zones. But how could they define territories of axons as well as their potential expansions? Fortunately—and this greatly contributed to designing the whole project—one interesting characteristic of the “dorsal hippocampus” helped them to establish points of reference common to all electron microscopic observable sections. It had indeed been demonstrated—and accepted—that the structure of the dorsal hippocampus looks like a grid, the dendrites of its cell bodies regularly intersecting axons indexed to different brain regions (Ramón y Cajal 1968). Therefore, if the brain researchers managed to produce electron microscopic observable slices of dorsal hippocampus extracted from similarly damaged rats’ brains (killed at different time intervals), the “natural” grid structure produced by the intersections of the dendrites of dorsal hippocampus’s body cells with axons indexed to different brain regions could constitute an initial empirical base for further measurements (Lynch 1985, 35–39). In other words, as it was certified that one specific part of the dorsal hippocampus contained cell bodies whose dendrites always intersected regularly with axons indexed to two different brain regions, which I call here α and β, it became possible to damage the β brain regions of all rats and then check if the axons indexed to α “sprouted” to infiltrate the territory of the axons previously indexed to β. But again, a new problem arose: how to go from specific electron microscopic views on slices to a panorama of many slices distributed over time? At the time of Lynch’s study, the easiest way to operate this translation was first to take analogical photographs of electron microscopic dorsal hippocampus displays. Brain scientists then had to develop these photographs in high definition and equip them with a coordinate system scaled according to the ultrastructural levels of observation (between 2,160 and 24,000 times, depending on the photographs). How did Lynch’s scientists concretely manage to equip these high-definition photographs? They pinned down the photographs on a cardboard sheet, hence creating a chronological montage of the microscopic displays. As Lynch put it, “these successions of photographs provided the visible configuration of brain ultrastructure that was addressed in the analytical phase of the study” (Lynch 1985, 38). But here again, it was not enough to measure an extension of axons indexed to α. Even though the dendrites of dorsal hippocampus’s cell bodies regularly intersected axons indexed to α and β, it remained necessary to affix a referential common to all photographs. How did the brain scientists do this? It is difficult here not to quote Lynch’s account:

在构建每个蒙太奇图时，都以以下方式进行分析：在照片表面铺上透明塑料片，并在塑料片表面上绘制线性刻度，该刻度沿垂直方向与照片柱状蒙太奇图的边缘平行。……为绘制的线绘制了“微米”刻度（根据照片的放大倍数计算），其中“零”点设置在一条水平线上，该水平线近似于颗粒细胞体层的排列。…… 沿此刻度的测量用于估计颗粒细胞树突从细胞体中出现并“向上”流动时沿“垂直”排列的线性距离。（Lynch 1985，38；斜体添加）

As each montage was constructed, it was analytically addressed in the following manner: a clear plastic sheet was laid over the surface of the photographs, and a linear scale was drawn over the surface of the sheet running in a vertical direction which paralleled the edge of the columnar montage of photographs. … A scale of “microns” (computed with reference to the magnificational power of the photographs) was plotted for the drawn-line, where the “zero” point was set at a horizontal line that approximated the alignment of the granule cell body layer. … Measurement along this scale was used to estimate linear distance along the “vertical” alignment of granule cell dendrites as they arose from the cell bodies and coursed “upward.” (Lynch 1985, 38; italics added)

平面线性距离与神经元及其轴突的潜在萌发相去甚远。然而，一旦将标准化大鼠背海马体小切片的放大照片装裱在纸板上，并配备在透明塑料片上绘制的线性刻度，其“零”点对应于每个切片的细胞体，这一古老的数学理论及其相关对象就会变得非常非常接近（Latour 1987，244）。实验室的实验环境及其所有产生“可对齐”铭文的仪器——标准化大鼠；大鼠背海马体的小切片，经过仔细清洗（和染色）；放大照片的蒙太奇；在透明塑料片上绘制的线性刻度——最终赋予大鼠背海马体与可以估计线性距离的图形相同的形式。在这一测量过程结束时，科学家甚至可以计算出按损伤后天数绘制的完整/死亡终端（轴突和树突之间的连接）的比例，从而从统计上证明轴突发芽的现象：“对这种扩张的测量表明，颗粒细胞树突区域下部 25% 的区域被重新占据，而这些区域以前是由（受损的）轴突层占据的”（Lynch 1985, 35）。

Flat linear distances are a priori far removed from neurons and the potential sprouting of their axons. Yet, once enlarged photographs of tiny little slices of standardized rats’ dorsal hippocampus are mounted on cardboard and equipped with a linear scale drawn on clear plastic sheets whose “zero” point corresponds to the cell body of each slice, this venerable mathematical theory and its correlated objects become very, very close (Latour 1987, 244). The experimental setting of the laboratory and all of its instruments producing “alignable” inscriptions—standardized rats; tiny, carefully washed (and stained) slices of rats’ dorsal hippocampus; montages of enlarged photographs; linear scales drawn on clear plastic sheets—end up conferring to rats’ dorsal hippocampus the same form as graphs on which linear distances can be estimated. At the end of this measurement process, ratios of intact/dead terminals—junctions between axons and dendrites—plotted in terms of days post the lesion could even be computed by the scientists, thus demonstrating statistically the phenomenon of axon sprouting: “Measurement of this expansion showed a consistent reoccupancy of the lower 25 per cent of the region of the granule cell dendrites formerly occupied by the [damaged] layer of axons” (Lynch 1985, 35).

再次，正如林奇所证明的那样，没有任何魔法介入；实验室实践使轴突和树突之间的关系变得可数学化。标准化的老鼠变成了背部海马体，小切片变成了放大的照片，纸板的蒙太奇变成了一个规则的几何空间，其占用情况随着时间的推移而演变。如果一些精致的数学事实——计算由完整终端逐渐占据的表面——确实有助于证明非数学现象（轴突发芽）的存在，那么这一事件需要一系列的翻译，以便将大脑湿润而笨重的生态与数学干燥而扁平的生态联系起来。

Again, as Lynch demonstrated, no magic intervened; laboratory practices made the relationships between axons and dendrites become mathematicable. Standardized rats became dorsal hippocampus, tiny slices became enlarged photographs, and a montage of cardboard became one regular geometrical space whose occupancy evolved through time. If some polished mathematical facts—computation of surfaces progressively occupied by intact terminals—did help demonstrate the existence of a nonmathematical phenomenon (axon sprouting), this event necessitated a succession of translations in order to connect the wet and bulky ecology of the brain with the dry and flat ecology of mathematics.

配方：定义

Formulating: A Definition

数学并不适用于世界。需要进行一系列的翻译才能将非数学实体与经过认证的数学事实联系起来。但是，在我们进行操作化练习的这个阶段，仍有一个问题：如果我们刚刚考虑过的大脑研究实验室的大鼠背海马和 Guillemin 实验室的大鼠垂体细胞培养物最终都被改造，以适应维持固化数学对象的网络（它们本身以前由逐渐成为经过认证的事实的主张描述，有时甚至是隐性无可争辩的知识的单句陈述），它们在路上不会失去许多属性吗？毕竟，从大脑丰富而复杂的区域来看，背海马变成了网格照片的修补蒙太奇；从丰富而复杂的细胞汤来看，大鼠垂体细胞培养物变成了一个简单的图表。为了使这两个实体都可数学化，它们必须经历重要的缩减。但这值得吗？什么理由使这种扁平化和干燥化？

Mathematics does not apply to the world. A cascade of translations is required to connect nonmathematical entities with certified mathematical facts. But at this point of our operationalization exercise, one question remains: if the rats’ dorsal hippocampus of the brain research laboratory we have just considered and the rat pituitary cell culture of Guillemin’s laboratory both end up being trans-formed in order to fit with the networks sustaining solidified mathematical objects (themselves formerly described by claims that progressively became certified facts and even, sometimes, single sentence statements part of tacit undisputable knowledge), do they not lose many properties on the road? After all, from a rich and complex region of the brain, the dorsal hippocampus becomes a tinkered montage of gridded photographs; from a rich and complex soup of cells, the rat pituitary cell culture becomes a simple graph. To make both entities mathematicable, they must endure important reductions. But is it worth it? What justifies such flattening and drying?

在这些特定情况下，这些简化的收益非常重要，因为数学家们以前在平坦实验室中定义的数学对象的属性逐渐“借给”垂体细胞培养物和背海马。首先，这两个实体都变得更容易处理。在从细胞汤到图表的转换过程之后，Guillemin 不再需要细胞汤了。他当然会将其保存起来以备将来的验证，但每当他需要查看或展示大鼠垂体细胞培养物时，他现在可以使用放射免疫分析打印的图表，该图表仅表达了汤的微小重要部分属性。林奇研究的大脑研究实验室也是如此：脑科学家现在不用处理海马体的微小切片，而是可以考虑网格照片。这种人体工程学优势的一个直接后果是，缩小的实体也变得更具共享性。虽然不可能通过电子邮件（在这些情况下是传真）发送湿漉漉、笨重的背部海马体，但在将它们转换成一系列照片后，位于世界另一端的值得信赖的脑科学家同事也能够仔细检查它们。将海马体转换成网格纸片使其能够投资扩展的（但广阔而脆弱的）通信网络。因此，这种缩小和扁平的海马体也变得更具可比性；如果位于世界另一端的脑科学家也设法对背部海马体进行类似的缩小，他们也许能够比较两组网格照片。 Guillemin 的图表也是如此：内分泌学家可以通过比较图表来比较细胞汤，而不是比较细胞汤，这是一个容易得多的事情。

In these specific situations, the gains of these reductions are important because the properties of the mathematical objects as formerly defined by mathematicians within their flat laboratories are progressively “lent” to the pituitary cell culture and the dorsal hippocampus. First, both entities become easier to handle. After the translation process from a cell soup to a graph, Guillemin does not need the cell soup anymore. He certainly conserves it for potential verifications, but whenever he needs to see or show the rat pituitary cell culture, he can now use the graph printed by the radioimmunoassay that expresses only the tiny important part of the soup’s properties. The same is true of the brain research laboratory studied by Lynch: instead of handling tiny slices of hippocampus, brain scientists can now consider gridded photographs. One direct consequence of this ergonomic gain is that the reduced entities become also more sharable. Although it is impossible to e-mail—or, in these cases, fax—wet and bulky dorsal hippocampus, after their translation into a succession of photographs, trustful brain scientist colleagues based on the other side of the world are also able to scrutinize them. Transforming the hippocampus into gridded pieces of paper allows it to invest extended—yet expansive and fragile—communication networks. Such a reduced and flattened hippocampus therefore also becomes more comparable; if the brain scientists based on the other side of the world also manage to operate similar reductions on the dorsal hippocampus, they may be able to compare both successions of gridded photographs. The same is also true of Guillemin’s graphs: instead of comparing cell soups, endocrinologists can compare graphs, a far easier endeavor.

减少实体并使其与经过认证的数学知识的平面网络相适应的另一个好处是，减少的实体变得更具可塑性；新的观点出现，反过来又提出了新的工具、测试和铭文。例如，当轴突和树突之间的活动连接变成均匀几何空间内的点时，数学家已经为这个几何空间定义的工具可用于进一步探测仍未定义的轴突发芽现象，从而产生新的铭文，这将有助于精确定义它。在这个几何空间内，可以进行新的测试，例如测量表面、计数终端和计算占用率。这些测试及其相关工具反过来会产生可读的铭文——这里是数字列表——这将有助于进一步描述所研究的现象。吉尔曼的大鼠垂体细胞培养也是如此：一旦复杂的生化反应变成随时间变化的离散值，所有通过这种图形形式可用的工具都可用于进一步探测细胞汤。图表的斜率是多少？生长激素减少的速度是多少？同样，扁平的简化形式使得能够使用新仪器并生成新的可读铭文，从而有助于描述新现象。

Another gain of reducing entities and making them fit with the flat network of certified mathematical knowledge is that reduced entities become much more malleable; new takes appear that, in turn, suggest new instruments, tests, and inscriptions. For example, when active junctions between axons and dendrites become points within a uniform geometrical space, the instruments already defined by mathematicians for this geometrical space can be used to further probe the still undefined phenomenon of axon sprouting, thus producing new inscriptions that will precisely help to define it. Within this geometrical space, new tests can be made, such as measuring surfaces, counting terminals, and calculating ratios of occupancy. These tests and their correlated instruments will, in turn, produce readable inscriptions—here, lists of numbers—that will help further characterize the phenomenon under scrutiny. The same is true of Guillemin’s rat pituitary cell culture: once complex biochemical reactions become discrete values varying through time, all the instruments that become available through this graphic form can be used to further probe the cell soup. What is the slope of the graph? What is the speed of the growth hormone’s decreasing? Again, a flat reduced form enables the use of new instruments and the production of new readable inscriptions that help with the characterization of a new phenomenon.

这引出了这些关键的简化过程的最后一个好处，也许是所有其他好处的结果：³⁹当一个实体与数学事实相兼容时，它也可以在书面声明中登记。将试图证明其物化存在。如果我们想要理解这些简化过程可能赋予未定义实体的全部额外力量，这一要素至关重要。如何将轴突纳入声称其具有发芽能力的文本中？如何将吉尔曼的新肽纳入证明其对生长激素作用减弱的论文中？将它们简化直到达到与经过认证的“平面”数学事实相同的形式，使它们成为将它们呈现给各自科学界的散文的指称。除了使轴突和肽更易于处理、更可共享、更可比较和更可塑之外，将它们简化以使其与数学事实的平面生态兼容，还可以将它们纳入讨论它们的文本中。具体化的对象“轴突发芽”不仅仅在论文中被描述，它还以平坦而干燥的形式存在于论文中，这正是使其数学化的原因（在这种情况下，根据Lynch [1985, 40–49] 的说法，它是一系列网格照片，其点“向上”移动）。同样，具体化的对象“生长抑素”不仅仅在论文中被描述，它还以总结其行为的图表的形式存在于论文中（Brazeau 等人，1973 年）。细心的读者可能已经注意到，我们现在已经从这项操作化练习的开始回到了原点，当时我们谈论的是书面的相对信念强度主张。实验室、实验、仪器和铭文的最终结果确实是试图吸引个人遵守的主张的表述。在这方面，我们现在应该能够更好地理解数学对象和事实的迷人力量；它们可能经历与其他科学事实相似的构造过程，但它们特定的平淡无奇的生态使它们与非数学对象和事实的形成相关。它们使未定义的实体更易于处理、更易于共享、更易于比较、更易于延展，并且更易于在它们确切帮助制定的主张中纳入。

This leads us to one last gain of these crucial reduction processes, perhaps the consequence of all the other gains:³⁹ when an entity is made compatible with mathematical facts, it also becomes enrollable within the written claim that will try to attest to its reified existence. This element is crucial if we want to understand the full additional strength these reduction processes may give to undefined entities. How indeed to include axons within a text claiming their ability to sprout? How to include Guillemin’s new peptide within a paper attesting to its decreasing effect on the growth hormone? Reducing them until they reach the same form as certified “flat” mathematical facts allows them to become the referents of the prose that presents them to their respective scientific communities. In addition to making both axons and peptide easier to handle, more shareable, more comparable, and more malleable, reducing them to make them compatible with the flat ecology of mathematical facts allows them to be included inside the texts that talk about them. The reified object “axon sprouting,” more than just being described in a paper, is also present within the paper in the flat and dry form that precisely allowed its mathematization (in this case, according to Lynch [1985, 40–49], as a succession of gridded photographs whose points move “upward”). Similarly, the reified object “somatostatin,” more than just being described in a paper, is also within the paper in the form of a graph summarizing its behavior (Brazeau et al. 1973). The attentive reader may have noticed that we have now come full circle from the beginning of this operationalization exercise where we were talking about written claims of relative conviction strengths. The end results of laboratories, experiments, instruments, and inscriptions are indeed the formulation of claims that try to attract the adherence of individuals. In this respect, we should now be in a position to better understand the fascinating power of mathematical objects and facts; they may go through construction processes that are similar to other scientific facts, but their particular flat and dry ecology makes them relevant for the formation of nonmathematical objects and facts. They make undefined entities easier to handle, more shareable, more comparable, more malleable, and more enrollable within claims they precisely help to formulate.

数学事实及其相关对象本身并不为它们有时遇到的转化实体提供额外的力量。相反，数学知识在其中部署的平面生态系统有时会为获得相同形式的实体提供优势。最后一个要素让我最终能够更技术性地定义公式化活动；在第三部分的剩余部分，我将公式化称为翻译未定义实体的经验过程直到它获得与已经定义的数学对象相同的形式。一个“扁平化”的实体与一个数学对象（之前必须在实验室中构建并通过主张呈现，而该主张的说服力使其成为一个完善的事实）之间的相遇，反过来将帮助科学家进一步描述实体的行为并在书面主张中呈现其具体化版本。就像任何科学主张（包括数学家提出的主张）一样，这个书面主张仍然必须克服出版、引用、标题和后人考验，才能最终成为认证的事实。一个圆圈已经画出；我们现在回到了原点。考虑到所有这些因素，现在是时候回到计算机科学的发展并参与人种学材料了。

It is not mathematical facts and their correlated objects that give, by themselves, some additional strength to the transformed entities they sometimes encounter. Rather, it is the flat ecology within which mathematical knowledge deploys itself that, sometimes, provides advantages to the entities that acquire the same form. This last element allows me to finally define the activity of formulating more technically; for the remainder of this part III, I shall call formulating the empirical process of translating an undefined entity until it acquires the same form as already defined mathematical object. The encounter between a “made-flat” entity and a mathematical object—that previously had to be constructed in a laboratory and presented in a claim whose conviction strength made it a polished fact—will, in turn, help scientists to further characterize the behavior of the entity and present its reified version in a written claim. Just as any scientific claim (including those formulated by mathematicians), this written claim will still have to overcome publication, citation, captation, and posterity trials to become, eventually, a certified fact. A circle has been drawn; we are now back to where we started. With all these elements in mind, it is high time to return to computer science in the making and engage with ethnographic materials.

笔记

Notes

1. 1.在这里，我的呈现风格和场景的运用深受拉图尔（1987）的启发。

2. 1.  Here, my style of presentation and use of scenes are greatly inspired by Latour (1987).

3. 2.我在这里遵循了 Rosental (2003) 的书。

4. 2.  I am following here Rosental’s (2003) book.

5. 3.我在此遵循 MacKenzie (1999) 的研究。

6. 3.  I am following here the work of MacKenzie (1999).

7. 4.这是从 2013 年 10 月至 2014 年 2 月的航海日志 1 中摘录的。

8. 4.  This is taken from Logbook 1, October 2013–February 2014.

9. 5.柏拉图主义哲学家可能通过区分apodeixis（严谨论证）和epideixis（修辞手法）开创了这种宏大叙事（Cassin 2014；Latour 1999）。根据 Leo Corry（1997）的说法，这种数学呈现方式最终导致了布尔巴基对数学真理的结构主义概念。关于这个主题，另请参阅 Lefebvre（2001，56–68）。有关宏大叙事的哲学探索，请参阅 Lyotard（1984）的经典著作。

10. 5.  With their distinction between apodeixis (rigorous demonstration) and epideixis (rhetorical maneuvering), Platonists philosophers may have initiated such grand narratives (Cassin 2014; Latour 1999). According to Leo Corry (1997), this way of presenting mathematics culminated with Bourbaki’s structuralist conception of mathematical truth. On this topic, see also Lefebvre (2001, 56–68). For a philosophical exploration of grand narratives, see the classic book by Lyotard (1984).

11. 6.然而，支持在 Facebook 或 Twitter 上发表的主张的“点赞”和“转发”有时可以作为重要的外部盟友。关于这一主题，请参阅 Ringelhan、Wollersheim 和 Welpe (2015)。

12. 6.  Yet “likes” and “retweets” that support claims published on Facebook or Twitter may, sometimes, work as significant external allies. On this topic, see Ringelhan, Wollersheim, and Welpe (2015).

13. 7 . 在 1878 年《美国数学杂志》 （AJM）创办之前，美国没有稳定的学术机构来发表数学研究成果（Kent 2008）。英国的情况略有不同：在《剑桥和都柏林数学杂志》的废墟上建立起来的《纯粹与应用数学季刊》（QJPAM）于 1855 年出版了第一期（Crilly 2004）。然而，对于肯普和希伍德的论文来说，他们期刊的编辑委员会——正如与今天的标准相比，当时的 AJM 成员人数较少（见前言）：1879 年AJM有 5 名成员（JJ Sylvester、WE Story、S. Newcomb、HA Newton 和 HA Rowland），1890 年QJPAM有 4 名成员（NM Ferrers、A. Cayley、JWL Glaisher 和 AR Forsyth）。

14. 7.  Before the 1878 foundation of the American Journal of Mathematics (AJM), there was no stable academic facility for the publication of mathematical research in the United States (Kent 2008). The situation in England was a bit different: built on the ashes of the Cambridge and Dublin Mathematical Journal, the Quarterly Journal of Pure and Applied Mathematics (QJPAM) published its first issue in 1855 (Crilly 2004). Yet for both Kempe’s and Heawood’s papers, the editorial boards of their journals—as indicated on their front matters—were rather small compared with today’s standards: five members for AJM in 1879 (J. J. Sylvester, W. E. Story, S. Newcomb, H. A. Newton, H. A. Rowland) and four members for QJPAM in 1890 (N. M. Ferrers, A. Cayley, J. W. L. Glaisher, A. R. Forsyth).

15. 8.根据美国人工智能协会（1993）的文件。

16. 8.  According to the document in American Association for Artificial Intelligence (1993).

17. 9.例如，请参阅《信息计量学杂志》。

18. 9.  See, for example, the Journal of Informetrics.

19. 10。简而言之，肯普将问题限定在平面上绘制的地图上，该地图至少包含一个邻国少于六个的地区，称为“国家”。然后，他可以将自己限制在五种情况，即邻国从一个到五个不等的国家。证明“四色性”对于有三个邻国的国家也适用显然不成问题。然而，为了证明对于有四个邻国的国家也适用，肯普使用了一种称为“肯普链”的论证（MacKenzie 1999，19–20）。该论证规定，对于有四个邻国 A、B、C、D 的国家 X，两个相对的邻国，比如 A 和 C，要么由一条连续的链（比如说红色和绿色国家）连接，要么不连接。如果它们由这样的红绿链连接，则 A 可以染成红色，而 C 可以染成绿色。但是，由于我们处理的是平面地图，X 的另外两个对立邻国——B 和 C——无法通过一条由蓝色和黄色国家组成的连续链条连接起来（不管怎样，这条链条确实被一个绿色或红色国家打断了）。因此，这两个对立的邻国可以涂成蓝色，而 X 可以涂成黄色。这样，对于有四个邻国的国家，四种着色性得以保留。肯普认为这种方法也适用于有五个邻国的国家。但是希伍德的图显示了这种方法失败的一个案例，其中 E 的红绿色区域（图 5.1中垂直交叉影线）与 B 的黄红色区域（水平交叉影线）相交，从而迫使两个国家都涂成红色。因此，X 必须涂成不同于红色、蓝色、黄色和绿色的颜色。在这种情况下，四种着色性无法保留。

20. 10.  In a nutshell, Kempe circumscribed the problem to maps drawn on a plane that contain at least one region called “country” with fewer than six neighbors. He could then limit himself to five cases, countries from one to up to five neighbors. Proving that “four colorability” is preserved for countries with three neighbors was, obviously, not a problem. Yet in order to prove it for countries with four neighbors, Kempe used an argument known as the “Kempe chains” (MacKenzie 1999, 19–20). This argument stipulates that for a country X with four neighbor countries A, B, C, D, two opposite neighbor countries, say A and C, are either joined by a continuous chain of, say, red and green countries, or they are not. If they are joined by such a red-green chain, A can be colored red and C can be colored green. But as we are dealing with a map drawn on a plane, the two other opposite neighbor countries of X—B and C—cannot be joined by a continuous chain of blue and yellow countries (one way or another, this chain is indeed interrupted by a green or red country). As a consequence, these two opposite neighbor countries can be colored blue and X can be colored yellow. Four colorability is thus preserved for countries with four neighbors. Kempe thought that this method also worked for countries with five neighbors. But Heawood’s figure shows a case of failure of this method where E’s red-green region (vertically cross-hatched in figure 5.1) intersects B’s yellow-red region (horizontally cross-hatched), thus forcing both countries to be colored red. Consequently, X has to be colored differently than red, blue, yellow, and green. In such a case, four colorability is not preserved.

21. 11.关于这一主题，请参阅Lefebvre (2001) 的著作。

22. 11.  On this topic, see the work of Lefebvre (2001).

23. 12 . 有关生命科学中的修辞习惯，请参阅 Latour and Woolgar (1986, 119–148) 和 Knorr-Cetina (1981, 94–130)。有关科学学科（不包括数学）之间的全面比较，请参阅 Penrose and Katz (2010)。

24. 12.  For rhetorical habits in the life sciences, see Latour and Woolgar (1986, 119–148) and Knorr-Cetina (1981, 94–130). For a thorough comparison among scientific disciplines—excluding mathematics—see Penrose and Katz (2010).

25. 13.尽管Serres(1995, 2002)做出了努力。

26. 13.  Despite the efforts made by Serres (1995, 2002).

27. 14。当然，当时没有科学机构；实验方案、同行见证以及后来的学术论文都是 17 世纪的产物（Shapin and Shaffer 1989）。然而，正如 Netz（2003，271–312）所表明的那样，写在蜡板和羊皮纸上的定理确实在少数（非常！）持怀疑态度的读者中流传。

28. 14.  There was, of course, no scientific institution at that time; experimental protocols, peer witnessing, and, later, academic papers are products of the seventeenth century (Shapin and Shaffer 1989). Yet, as Netz (2003, 271–312) showed, theorems written on wax tablets and parchments did circulate among a restricted audience of (very!) skeptical readers.

29. 15 . 这至少是 Netz (2003, 271–304) 的假设，并得到了 Lloyd (1990, 2005) 的研究支持。正如拉图尔所总结的那样：“正是因为公共生活“希腊太具侵略性，太好辩，太没有定论，因此，‘自学成才的高度专业化网络’发明了另一种方式来结束无休止的讨论，这具有如此诱人的吸引力”(Latour 2008, 449)。

30. 15.  This is at least Netz’s (2003, 271–304) hypothesis, supported by the work of Lloyd (1990, 2005). As Latour summarized it: “It is precisely because the public life in Greece was so invasive, so polemical, so inconclusive, that the invention, by ‘highly specialized networks of autodidacts’, of another way to bring an endless discussion to a close took such a tantalizing aspect” (Latour 2008, 449).

31. 16 . 令人惊讶的是，这种谨慎而高度专业化的信念方法，掌握在一群自学成才的边缘群体手中，他们非常小心地坚持形式，但很快被柏拉图“借用”并扩展到内容，目的之一是让智者们闭嘴。至少这是卡辛（2014）、拉图尔（1999b，216-235）和 Netz（2004，275-282）提出的论点。

32. 16.  So surprising that this careful and highly specialized method of conviction mastered by a peripheral community of autodidacts who took great care to stick to forms was soon “borrowed” by Plato and extended to content in order to, among other things, silence the Sophists. This is at least the argument made by Cassin (2014), Latour (1999b, 216–235), and Netz (2004, 275–282).

33. 17 . 亚里士多德似乎是最早编纂几何文献并系统化其逻辑论证的人之一（Bobzien 2002）。在古代晚期，Eutocius 等评论家注释了许多几何著作并汇编了它们的主要结果，以便于进行系统比较（Netz 1998）。根据 Netz (2004) 的说法，这些标准化几何汇编的集合进一步帮助了 al-Kwarizmi 和 Khayyam 等伊斯兰数学家构建代数语言。

34. 17.  Aristotle seems to be one of the first to compile geometrical texts and systematize their logical arguments (Bobzien 2002). During late antiquity, commentators such as Eutocius annotated many geometrical works and compiled their main results to facilitate their systematic comparisons (Netz 1998). According to Netz (2004), these collections of standardized geometrical compilations further helped Islamic mathematicians such as al-Kwarizmi and Khayyam to constitute the algebraic language.

35. 18.在十九世纪末所谓的数学基础危机期间，以大卫·希伯特为首的形式主义学派试图将数学的基础建立在逻辑原理上（Corry 1997）。这导致了著名的失败，例如罗素和怀特海的三卷《数学原理》（Whitehead and Russell 1910、1911、1913）。多亏了 Netz 的语言学工作，我们现在更好地理解了为什么这样的努力会失败：正是数学实践——精心索引到希腊小句子的字母图表——导致了逻辑规则的制定，而不是相反。

36. 18.  During the late nineteenth century’s so-called crisis of foundations in mathematics, the formalist school—headed by David Hibert—tried to establish the foundations of mathematics on logical principles (Corry 1997). This led to famous failures such as Russell and Whitehead’s three volumes of Principia Mathematica (Whitehead and Russell 1910, 1911, 1913). Thanks to the philological work of Netz, we now better understand why such an endeavor has failed: it was the very practice of mathematics—lettered diagrams carefully indexed to small Greek sentences—that led to the formulation of the rules of logic and not the other way round.

37. 19 . 在一定程度上，Lefebvre (2001) 和 Mialet (2012) 除外。看来，拉图尔的评论仍然是正确的：很少有学者有勇气对数学进行仔细的人类学研究 (Latour 1987, 246)。

38. 19.  Except, to a certain extent, Lefebvre (2001) and Mialet (2012). It seems then that Latour’s remark remains true: few scholars have had the courage to do a careful anthropological study of mathematics (Latour 1987, 246).

39. 20.这取自Latour（1987，第2章）和Wade（1981，第13章）。

40. 20.  This is taken from Latour (1987, chapter 2) and Wade (1981, chapter 13).

41. 21.摘自Pickering and Stephanides (1992) 以及 Hankins (1980, 280–312)。

42. 21.  This is taken from Pickering and Stephanides (1992) and Hankins (1980, 280–312).

43. 22 . 肽是由氨基酸链组成的化学元素。众所周知，它们与激素密切相关。由于氨基酸种类繁多（人类有 20 种），因此可能存在由 2 到 50 种氨基酸组合而成的数十亿种不同的肽。值得注意的是，在 1972 年 Guillemin 进行实验时，肽已经可以在设备齐全的实验室中组装和探测。

44. 22.  Very schematically, peptides are chemical elements made of chains of amino acids. They are known for interacting intimately with hormones. As there are many different amino acids (twenty for the case of humans), there exists—potentially—billions of different peptides made of combinations of two to fifty amino-acids. It is important to note that in 1972, at the time of Guillemin’s experiment, peptides could already be assembled—and probed—within well-equipped laboratories.

45. 23.在汉密尔顿时代，复数（所谓的荒谬数量，例如负数的平方根）的标准代数符号是x + iy，其中i ² = –1， x和y是实数。早期复数代数的这些进步对几何学家来说是个问题：如果正实数可以被视为可测量的数量，那么负实数及其平方根就很难表示为平面上的形状。克服这一僵局的一种方法是将x和y视为终止于原点的线段端点的坐标。因此，“平面的x轴测量表示为这种线段的给定复数的实部，y轴测量虚部，即代数表达式中乘以i 的部分”（Pickering 和 Stephanides 1992，145）。通过这种复数的可视化，像汉密尔顿这样的代数几何学家可以将线段上的复杂几何运算与方程上的复杂代数运算联系起来。就这样，几何学和复代数之间的桥梁就建立起来了。然而，几何学并不局限于平面：如果二维线段 [0, x + iy ] 可以表示一个复数，那么三维线段 [0, x + iy + jz ] 没有理由不能表示另一个复数。表征这种线段的行为是汉密尔顿实验的既定目标。

46. 23.  At the time of Hamilton, the standard algebraic notation for a complex number—so-called absurd quantities such as square roots of negative numbers—was x + iy, where i² = –1 and x and y are real numbers. These advances in early complex algebra were problematic to geometers: if positive real numbers could be considered measurable quantities, negative real numbers and their square roots were difficult to represent as shapes on a plane. A way to overcome this impasse was to consider x and y as coordinates of the end point of a segment terminating at the origin. Therefore, “the x-axis of the plane measured the real component of a given complex number represented as such a line segment, and the y axis the imaginary part, the part multiplied by i in the algebraic expression” (Pickering and Stephanides 1992, 145). With this visualization of complex numbers, algebraic geometers such as Hamilton could relate complex geometrical operations on segments and complex algebraic operations on equations. A bridge between geometry and complex algebra was thus built. Yet geometry is not confined to planes: if a two-dimensional segment [0, x + iy] can represent a complex number, there is a priori no reason why a three-dimensional segment [0, x + iy + jz] could not represent another complex number. Characterizing the behavior of such a segment was the stated goal of Hamilton’s experiment.

47. 24 . 汉密尔顿对复数理论与几何学之间关系的探究并非纯粹的探索性努力。正如 Pickering 和 Stephanides 所指出的，“希望构建三维空间中线段变换的代数复制，从而开发出一种新的、可能有用的代数系统，适用于三维几何计算”（Pickering and Stephanides 1992，146）。

48. 24.  Hamilton’s inquiry into the relationships between complex number theory and geometry was not a purely exploratory endeavor. As Pickering and Stephanides noted, “the hope was to construct an algebraic replica of transformations of line segments in three-dimensional space and this to develop a new and possibly useful algebraic system appropriate to calculations in three-dimensional geometry” (Pickering and Stephanides 1992, 146).

49. 25与汉密尔顿相反，古希腊几何学家只能用简短但繁琐的希腊语句来引用字母图表（Netz 2003，127–167）。除了希腊几何学家对微分的重视外，缺乏像代数这样的浓缩语言（这恰恰需要编纂几何著作集才能构成）（Netz 1998）可能也限制了古希腊几何命题的范围（Netz 2004，11–54）。

50. 25.  Contrary to Hamilton, ancient Greek geometers could only refer to their lettered diagrams with short but still cumbersome Greek sentences (Netz 2003, 127–167). Along with Greek geometers’ emphasis on differentiation, the absence of a condensed language such as algebra—that precisely required compiled collections of geometrical works in order to be constituted (Netz 1998)—may have participated in limiting the scope of ancient Greek geometrical propositions (Netz 2004, 11–54).

51. 26。关于这些工具，值得一提的是，我们在这里重拾了上一节讨论的内容：所有这些工具（也许除了非交换代数）都是黑箱中经过打磨的事实，最初都是书面声明。大鼠垂体细胞培养、代数符号、放射免疫测定、坐标空间，甚至毕达哥拉斯定理都必须经过试验才能获得说服力，成为既定的、经过认证的事实。

52. 26.  Regarding these instruments, it is worth mentioning that here we retrieve what we were discussing about in the last section: all of them—except, perhaps, noncommutative algebra—are blackboxed polished facts that were, initially, written claims. Rat pituitary cell cultures, algebraic notations, radioimmunoassays, coordinate spaces and even Pythagoras’s theorem all had to overcome trials in order to gain conviction strength and become established, certified facts.

53. 27.数学实验室的这种拓扑特性可能是为什么它们很少成为民族志研究场所的原因 (Latour 2008, 444)。

54. 27.  This topological characteristic of mathematical laboratories may be a reason why they have rarely been sites for ethnographic inquiries (Latour 2008, 444).

55. 28。当然，正如我们在第 4 章中看到的那样，如果没有之前制作这些铭文所需的整个系列，这些铭文就毫无意义。只有通过将“最终”铭文与之前的铭文对齐，从而形成一个参考链，Guillemin 才能提供有关他的肽的信息（Latour 2013，第 3 章）。

56. 28.  Of course, as we saw in chapter 4, such inscriptions are meaningless without the whole series of inscriptions previously required to produce them. It is only by aligning the “final” inscriptions to former ones, thus creating a chain of reference, that Guillemin can produce information about his peptide (Latour 2013, chapter 3).

57. 29。在这里，我们找回了在第 3 章和第 4 章中已经遇到过的东西：对齐铭文的“认知”实践。就像 DF 在他的计算机终端前一样，Brazeau、Guillemin 和 Hamilton 从未停止掌握他们从实验中获得的铭文。反过来，这些铭文可以被视为建议进一步行动的行动。

58. 29.  Here we retrieve something we already encountered in chapters 3 and 4: the “cognitive” practice of aligning inscriptions. Just as DF in front of his computer terminal, Brazeau, Guillemin, and Hamilton never stop grasping inscriptions they acquire from experiments. These inscriptions can, in turn, be considered takes suggesting further actions.

59. 30.同样，这取自 Latour (1987, 第 2 章) 和 Wade (1981, 第 13 章)。

60. 30.  Again, this is taken from Latour (1987, chapter 2) and Wade (1981, chapter 13).

61. 31.同样，这一观点摘自Pickering and Stephanides (1992) 以及 Hankins (1980, 280–312)。

62. 31.  Again, this is taken from Pickering and Stephanides (1992) and Hankins (1980, 280–312).

63. 32。Brazeau和 Guillemin在《科学》杂志上发表了他们的研究结果（Brazeau 等，1973）。1843 年 11 月，Hamilton 在爱尔兰皇家科学院发表了他的研究成果，随后在《伦敦、爱丁堡和都柏林哲学杂志和科学杂志》上发表了关于四元数的论文（Hamilton，1844）。关于四元数，需要注意的一点是，Hamilton 这样命名四元数之后，仍然需要定义复数k ²、 ik、 kj和i ²才能完成他的系统。根据 Hamilton 在 1865 年写的一封信，这个问题的解——众所周知的i ² = j ² = k ² = ijk = − 1——是在他沿着都柏林皇家运河散步时突然出现的。如果这一时刻确实很重要，那么称之为“四元数的发现”则是错误的（Buchman，2009）。正如 Pickering 和 Stephanides (1992) 所指出的，在为虚数乘积赋予值之前，四元数就已经被定义为对象了。事实上，与定义这些乘积值的问题所需的实验工作相比，都柏林皇家运河上发生的事情显得微不足道。

64. 32.  Brazeau and Guillemin published their results in Science (Brazeau et al. 1973). After having presented his results at the Royal Irish Academy in November 1843, Hamilton published a paper on quaternions in The London, Edinburg and Dublin Philosophical Magazine and Journal of Science (Hamilton 1844). An important thing to note about quaternions is that after Hamilton named them that way, he still had to define the complex quantities k², ik, kj, and i² in order to complete his system. According to a letter Hamilton wrote in 1865, the solution to this problem—the well-known i² = j² = k² = ijk = −1—appeared to him as he was walking along the Royal Canal in Dublin. If this moment was indubitably important, it would be erroneous to call it “the discovery of quaternions” (Buchman 2009). As shown by Pickering and Stephanides (1992), quaternions were already defined as objects before the attribution of values to the imaginary quantities’ products. In fact, when compared with the experimental work required to define the problem of these products’ values, what happened on Dublin’s Royal Canal appears relatively minor.

65. 33.这是重要数学家传记中经常出现的问题；因为他们倾向于用自然来解释伟大的成就，他们往往忽略了塑造“发现”物体所需的许多仪器和铭文。因此，伟大数学家的传记往往——但并非总是如此（参见令人惊叹的连载漫画《Logicomix》 [Doxi à dis et al. 2010]）——讲述自然选择的孤独天才的不切实际的故事。

66. 33.  This is the recurrent problem of biographies of important mathematicians; as they tend to use nature to explain great achievements, they often ignore the many instruments and inscriptions that were needed to shape the “discovered” objects. Biographies of great mathematicians are thus often—yet not always (see the amazing comic strip Logicomix [Doxiàdis et al. 2010])—unrealistic stories of solitary geniuses chosen by nature.

67. 34。接受自然的双重性——既定争议的结果，也是无争议事实的回顾性原因——为我们提供了一个全新的视角来审视数学哲学中柏拉图主义和直觉主义之间的古典对立。柏拉图主义（认为数学对象来自外部理念世界）和直觉主义（认为数学对象来自人类意识的内部世界）的奇怪之处似乎来自于它们共同的出发点：他们都考虑经过认证的无争议的数学事实。然而，一旦人们考虑到数学中的争议——即正在形成的数学——来自上方的自然（理念的外部世界）或来自下方的自然（人类意识的内部世界）就不能再被视为资源，因为这两者正是争论中所关心的。然而，值得注意的是，这两种对立的数学起源的非经验主义观念导致了关于数学实践的重要表现分歧，特别是通过接受或拒绝排中律。关于这个有趣的话题，请参阅 Rotman (2006) 和 Corry (1997)。

68. 34.  Accepting the dual aspect of nature—the consequence of settled controversies as well as the retrospective cause of noncontroversial facts—provides a fresh new look at the classical opposition between Platonism and Intuitionism in the philosophy of mathematics. It seems indeed that the oddity of both Platonism—for which mathematical objects come from the outer world of ideas—and Intuitionism—for which mathematical objects come from the inner world of human consciousness—comes from their shared starting point: they both consider certified noncontroversial mathematical facts. Yet as soon as one accounts for controversies in mathematics—that is, mathematics in the making—nature from above (the outer-world of ideas) or nature from below (the inner-world of human consciousness) cannot be considered resources anymore as both are precisely what is at stake during the controversies. It is interesting to note, however, that both antagonist unempirical conceptions of the origin of mathematics led to important performative disagreements about the practice of mathematics, notably through the acceptance, or refusal, of the law of excluded middle. On this fascinating topic, see Rotman (2006) and Corry (1997).

69. 35.根据 Netz (2004, 181–186) 的说法，对古代数学文本中差异性和原创性的不断探索，产生了个体类似问题的证明表述不同。简而言之，希腊几何学家对系统不感兴趣；他们对具有特定“光环”的真实证明感兴趣（Netz 2004，58–63）。

70. 35.  According to Netz (2004, 181–186), the constant search for differentiation and originality in ancient mathematical texts had the effect of multiplying individual proofs of similar problems stated differently. In short, Greek geometers were not interested in systems; they were interested in authentic proofs with a specific “aura” (Netz 2004, 58–63).

71. 36。Netz 认为，古代数学文本的论战动态阻碍了希腊数学家使他们的作品、演示和问题规范化。正如他所指出的：“我们迄今为止所看到的策略——希腊数学家试图将其作品与其背景隔离开来——现在被视为既审慎又有效。它之所以审慎，是因为它是一种提前保护作品不被卷入你无法控制的文本间论战的方式。它之所以有效，是因为它使你的作品熠熠生辉，仿佛超越了论战。当希腊数学家通过明确的介绍或隐含地通过问题的数学陈述为他们的文本奠定基础时，他们旨在将一切都一笔勾销：使新命题尽可能地作为一个独特的事件出现——即手头问题的第一个真正的解决方案” (Netz 2004, 62–63)。

72. 36.  Netz suggests that the polemical dynamics of ancient mathematical texts prevented Greek mathematicians from normalizing their works, demonstrations, and problems. As he noted: “The strategy we have seen so far—of the Greek mathematician trying to isolate his work from its context—is seen now as both prudent and effective. It is prudent because it is a way of protecting the work, in advance, from being dragged into inter-textual polemics over which you do not have control. And it is effective because it makes your work shine, as if beyond polemic. When Greek mathematicians set out the ground for their text, by an explicit introduction or, implicitly, by the mathematical statement of the problem, what they aim to do is to wipe the slate clean: to make the new proposition appear, as far as possible, as a sui generis event—the first genuine solution of the problem at hand” (Netz 2004, 62–63).

73. 37.在一定程度上，正如我们将在第 6 章中看到的，Wolfram Mathematica 和 Matlab 等数学软件可以被视为经过完善、编译和标准化的数学认证知识的存储库。

74. 37.  To a certain extent, as we will shall see in chapter 6, mathematical software such as Wolfram Mathematica and Matlab can be considered repositories of polished, compiled, and standardized mathematical certified knowledge.

75. 38.神经元细胞由三部分组成。首先是“树突”：允许神经元接收电化学信号的结构。然后是“细胞体”：神经元的球形部分，包含细胞核并对信号作出反应。最后是“轴突”：延伸的细胞膜，向其他树突发送信息。

76. 38.  Very schematically, a neuron cell is made of three parts. There is first the “dendrite”: the structure that allows a neuron to receive an electro-chemical signal. There is then the “cell body”: the spherical part of the neuron that contains the nucleus of the cell and reacts to the signal. There is finally the “axon”: the extended cell membrane that sends information to other dendrites.

77. 39.值得注意的是，伴随归约过程而来的不可避免的损失可用于批评这些归约的结果。这正是我在第 3 章处理思维的计算隐喻时所做的。我利用了一些归约没有考虑到的东西来批评这些归约的结果。

78. 39.  It is important to note that the inevitable losses that go along with reduction processes can be used to criticize the products of these reductions. This is exactly what I did in chapter 3 when I was dealing with the computational metaphor of the mind. I used what some reductions did not take into account in order to criticize the product of these reductions.

实证材料的呈现

Presentation of the Empirical Materials

本案例研究取自我们在第 2 章中已经遇到的显著性检测项目。为了让读者重温一下，这个显著性检测项目包括两名博士生和一名博士后——BJ、GY 和 CL——我将继续称他们为一个实体：“小组”。简而言之，小组提出该项目的论点是，如果显著性检测算法能够检测、分割和评估复杂数字照片中显著物体和人脸的不同重要性，那么图像处理中的显著性检测可能会在工业上变得更有趣。显著性问题的这种新问题化要求构建一个新的地面实况数据库，收集未标记的复杂数字图像及其手动标记的对应物，即“目标”。新的地面实况对于小组算法的形成至关重要，因为这个数据库实质上确定了要通过计算解决的问题的条件。为了有效地塑造其算法，小组将其新的地面实况数据库分为两组：训练集和评估集。训练集用于研究输入数据与目标之间的关系。一旦这些关系被定义并表达在计算模型中，该小组就将此模型转换为机器可读的指令编号列表，从而组装出一个真正的计算机程序。然后可以通过标准统计方法在基本事实评估集上评估该程序的性能。新的基本事实数据库、计算模型的原理以及相关计算机程序的处理性能后来在一篇学术论文中提出，这篇论文被一个重要的图像处理会议的委员会拒绝了。然而一年后，这篇文章的修订版在一个较小的会议上获得了“最佳短文奖”。

This case study is taken from the saliency-detection project we already encountered in chapter 2. Just to refresh the memory of the reader, this saliency-detection project included two PhD students and a postdoc—BJ, GY, and CL—that I shall keep on referring to as a single entity: “the Group.” In a nutshell, the Group’s argument that framed the project was that saliency detection in image processing may become industrially more interesting if saliency-detection algorithms could detect, segment, and evaluate the varying importance of salient objects and human faces within complex digital photographs. This new problematization of the saliency problem called for the construction of a new ground-truth database gathering unlabeled complex digital images and their manually labeled counterparts, the “targets.” The new ground truth was central to the formation of the Group’s algorithm as this database materially established the terms of the problem to be solved computationally. To effectively shape its algorithm, the Group divided its new ground-truth database into two sets: a training set and an evaluation set. The training set was used to study the relationships between input-data and their targets. Once these relationships were defined and expressed in a computational model, the Group translated this model into numbered lists of machine-readable instructions, thus assembling a genuine computer program. The performances of this program could then be evaluated on the evaluation set of the ground truth by means of standard statistical measures. The new ground-truth database, the principles of the computational model, and the processing performances of the correlated computer program were later presented in an academic paper that was rejected by the committee of an important conference in image processing. Yet one year later, a revised version of the article won the “Best Short Paper Award” at a smaller conference.

在以下章节中，我将主要关注训练集和导致制定输入图像与其目标之间关系的实践，然后将其转换为代码行。由于该小组的新地面实况的目标非常复杂，我将专注于目标的一个组成部分：检测到的和分割的人脸的相对重要性值（见图6.1）。我的目标是解释导致表征一种自动计算检测到的人脸的相对重要性值的方法的制定实践，从而检索地面实况目标的一小部分。会计因为这些实践将使我能够将第三部分与第一部分（地面实况）和第二部分（编程）联系起来。本案例研究还将作为触及现在广泛讨论的机器学习和人工智能主题的垫脚石。

In the following sections, I will mainly focus on the training set and the practices that led to the formulation of the relationships between input-images and their targets that was then translated into lines of code. As the targets of the Group’s new ground truth were quite complex, I will focus exclusively on one of the targets’ component: the relative importance values of the detected and segmented faces (see figure 6.1). My goal is to account for the formulating practices that led to the characterization of a way to automatically calculate the relative importance values of detected faces, thus retrieving one—small—part of the ground truth’s targets. Accounting for these practices will allow me to link this part III with part I (ground-truthing) and part II (programming). This case study will also serve as stepping stone to touch on the now widely discussed topics of machine learning and artificial intelligence.

为了更好地理解导致定义面部重要性计算模型的实践，我们必须仔细研究该组的训练集及其数据的逐步重组。然而，作为 Matlab由于训练集相当混乱（见图6.2），我无法基于“真实”屏幕截图进行分析。就像在第 4 章中我讲解编程实践时一样，我必须简化组的训练集并仅保留与当前分析相关的元素。因此，组训练集的简化版本将如表 6.1所示。由于我们将跟踪一系列翻译，因此组训练集的第一个翻译将算作一，第二个翻译算作二，依此类推。训练集的初始形式将算作翻译 0。

To better understand the practices that lead to the definition of a computational model for face importance, we will have to closely examine the Group’s training set and the progressive reorganization of its data. Yet, as a Matlab training set is quite confusing (see figure 6.2), I will not be able to base my analysis on “real” screenshots. Just like in chapter 4 when I was accounting for programming practices, I will have to simplify the Group’s training set and retain only the elements that are relevant for the present analysis. The simplified version of the Group’s training set will thus be presented as in table 6.1. As we are going to follow a succession of translations, the first translation of the Group’s training set will be counted as one, the second translation as two, and so on. The initial form of the training set will be counted as translation 0.